InvokeAI/invokeai/backend/training/textual_inversion_training.py

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

935 lines
34 KiB
Python
Raw Normal View History

use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# This code was copied from
# https://github.com/huggingface/diffusers/blob/main/examples/textual_inversion/textual_inversion.py
# on January 2, 2023
# and modified slightly by Lincoln Stein (@lstein) to work with InvokeAI
2023-01-26 20:10:16 +00:00
"""
This is the backend to "textual_inversion.py"
"""
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
import logging
import math
import os
import random
from pathlib import Path
from typing import Optional
2023-01-26 20:10:16 +00:00
import datasets
import diffusers
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
import numpy as np
2023-01-26 20:10:16 +00:00
import PIL
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
2023-08-18 15:13:28 +00:00
from accelerate.utils import ProjectConfiguration, set_seed
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
from diffusers.optimization import get_scheduler
from diffusers.utils import check_min_version
from diffusers.utils.import_utils import is_xformers_available
from huggingface_hub import HfFolder, Repository, whoami
# TODO: remove and import from diffusers.utils when the new version of diffusers is released
from packaging import version
from PIL import Image
2023-01-26 20:10:16 +00:00
from torch.utils.data import Dataset
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer
2023-01-26 20:10:16 +00:00
# invokeai stuff
from invokeai.app.services.config import InvokeAIAppConfig, PagingArgumentParser
2023-10-23 22:48:14 +00:00
from invokeai.app.services.model_manager import ModelManagerService
from invokeai.backend.model_management.models import SubModelType
2023-01-26 20:10:16 +00:00
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"):
PIL_INTERPOLATION = {
"linear": PIL.Image.Resampling.BILINEAR,
"bilinear": PIL.Image.Resampling.BILINEAR,
"bicubic": PIL.Image.Resampling.BICUBIC,
"lanczos": PIL.Image.Resampling.LANCZOS,
"nearest": PIL.Image.Resampling.NEAREST,
}
else:
PIL_INTERPOLATION = {
"linear": PIL.Image.LINEAR,
"bilinear": PIL.Image.BILINEAR,
"bicubic": PIL.Image.BICUBIC,
"lanczos": PIL.Image.LANCZOS,
"nearest": PIL.Image.NEAREST,
}
# ------------------------------------------------------------------------------
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.10.0.dev0")
logger = get_logger(__name__)
2023-01-26 20:10:16 +00:00
def save_progress(text_encoder, placeholder_token_id, accelerator, placeholder_token, save_path):
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
logger.info("Saving embeddings")
2023-01-26 20:10:16 +00:00
learned_embeds = accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[placeholder_token_id]
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
learned_embeds_dict = {placeholder_token: learned_embeds.detach().cpu()}
torch.save(learned_embeds_dict, save_path)
2023-01-26 20:10:16 +00:00
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
def parse_args():
config = InvokeAIAppConfig.get_config()
parser = PagingArgumentParser(description="Textual inversion training")
2023-01-26 20:10:16 +00:00
general_group = parser.add_argument_group("General")
model_group = parser.add_argument_group("Models and Paths")
image_group = parser.add_argument_group("Training Image Location and Options")
trigger_group = parser.add_argument_group("Trigger Token")
training_group = parser.add_argument_group("Training Parameters")
checkpointing_group = parser.add_argument_group("Checkpointing and Resume")
integration_group = parser.add_argument_group("Integration")
general_group.add_argument(
2023-01-26 20:10:16 +00:00
"--front_end",
"--gui",
dest="front_end",
action="store_true",
default=False,
2023-01-26 20:10:16 +00:00
help="Activate the text-based graphical front end for collecting parameters. Aside from --root_dir, other parameters will be ignored.",
)
general_group.add_argument(
2023-01-26 20:10:16 +00:00
"--root_dir",
"--root",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
type=Path,
default=config.root,
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
help="Path to the invokeai runtime directory",
)
general_group.add_argument(
"--logging_dir",
type=Path,
default="logs",
help=(
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
),
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
general_group.add_argument(
"--output_dir",
type=Path,
default=f"{config.root}/text-inversion-model",
help="The output directory where the model predictions and checkpoints will be written.",
)
model_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--model",
type=str,
default="sd-1/main/stable-diffusion-v1-5",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
help="Name of the diffusers model to train against, as defined in configs/models.yaml.",
)
model_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--revision",
type=str,
default=None,
required=False,
help="Revision of pretrained model identifier from huggingface.co/models.",
)
2023-01-26 20:10:16 +00:00
model_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--tokenizer_name",
type=str,
default=None,
help="Pretrained tokenizer name or path if not the same as model_name",
)
image_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--train_data_dir",
type=Path,
default=None,
2023-01-26 20:10:16 +00:00
help="A folder containing the training data.",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
image_group.add_argument(
"--resolution",
type=int,
default=512,
help=(
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
" resolution"
),
)
image_group.add_argument(
2023-01-26 20:10:16 +00:00
"--center_crop",
action="store_true",
help="Whether to center crop images before resizing to resolution",
)
trigger_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--placeholder_token",
"--trigger_term",
2023-01-26 20:10:16 +00:00
dest="placeholder_token",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
type=str,
default=None,
2023-01-26 20:10:16 +00:00
help='A token to use as a placeholder for the concept. This token will trigger the concept when included in the prompt as "<trigger>".',
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
trigger_group.add_argument(
"--learnable_property",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
type=str,
2023-01-26 20:10:16 +00:00
choices=["object", "style"],
default="object",
2023-01-26 20:10:16 +00:00
help="Choose between 'object' and 'style'",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
trigger_group.add_argument(
"--initializer_token",
type=str,
2023-01-26 20:10:16 +00:00
default="*",
help="A symbol to use as the initializer word.",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
checkpointing_group.add_argument(
"--checkpointing_steps",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
type=int,
default=500,
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
help=(
"Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming"
" training using `--resume_from_checkpoint`."
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
),
)
checkpointing_group.add_argument(
"--resume_from_checkpoint",
type=Path,
default=None,
help=(
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
),
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
checkpointing_group.add_argument(
"--save_steps",
type=int,
default=500,
help="Save learned_embeds.bin every X updates steps.",
)
training_group.add_argument(
2023-01-26 20:10:16 +00:00
"--repeats",
type=int,
default=100,
help="How many times to repeat the training data.",
)
training_group.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
training_group.add_argument(
"--train_batch_size",
type=int,
default=16,
help="Batch size (per device) for the training dataloader.",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
training_group.add_argument("--num_train_epochs", type=int, default=100)
training_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--max_train_steps",
type=int,
default=5000,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
training_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
training_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--gradient_checkpointing",
action="store_true",
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
)
training_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--learning_rate",
type=float,
default=1e-4,
help="Initial learning rate (after the potential warmup period) to use.",
)
training_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--scale_lr",
action="store_true",
default=True,
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
)
training_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--lr_scheduler",
type=str,
default="constant",
help=(
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
' "constant", "constant_with_warmup"]'
),
)
training_group.add_argument(
2023-01-26 20:10:16 +00:00
"--lr_warmup_steps",
type=int,
default=500,
help="Number of steps for the warmup in the lr scheduler.",
)
training_group.add_argument(
"--adam_beta1",
type=float,
default=0.9,
help="The beta1 parameter for the Adam optimizer.",
)
training_group.add_argument(
"--adam_beta2",
type=float,
default=0.999,
help="The beta2 parameter for the Adam optimizer.",
)
training_group.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
training_group.add_argument(
"--adam_epsilon",
type=float,
default=1e-08,
help="Epsilon value for the Adam optimizer",
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
training_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--mixed_precision",
type=str,
default="no",
choices=["no", "fp16", "bf16"],
help=(
"Whether to use mixed precision. Choose"
"between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
"and an Nvidia Ampere GPU."
),
)
training_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--allow_tf32",
action="store_true",
help=(
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
),
)
2023-01-26 20:10:16 +00:00
training_group.add_argument(
"--local_rank",
type=int,
default=-1,
help="For distributed training: local_rank",
)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
parser.add_argument(
2023-01-26 20:10:16 +00:00
"--enable_xformers_memory_efficient_attention",
action="store_true",
help="Whether or not to use xformers.",
)
integration_group.add_argument(
"--only_save_embeds",
action="store_true",
default=False,
help="Save only the embeddings for the new concept.",
)
integration_group.add_argument(
"--hub_model_id",
type=str,
default=None,
help="The name of the repository to keep in sync with the local `output_dir`.",
)
integration_group.add_argument(
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
"--report_to",
type=str,
default="tensorboard",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
),
)
2023-01-26 20:10:16 +00:00
integration_group.add_argument(
"--push_to_hub",
action="store_true",
help="Whether or not to push the model to the Hub.",
)
integration_group.add_argument(
"--hub_token",
type=str,
default=None,
help="The token to use to push to the Model Hub.",
)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
args = parser.parse_args()
return args
imagenet_templates_small = [
"a photo of a {}",
"a rendering of a {}",
"a cropped photo of the {}",
"the photo of a {}",
"a photo of a clean {}",
"a photo of a dirty {}",
"a dark photo of the {}",
"a photo of my {}",
"a photo of the cool {}",
"a close-up photo of a {}",
"a bright photo of the {}",
"a cropped photo of a {}",
"a photo of the {}",
"a good photo of the {}",
"a photo of one {}",
"a close-up photo of the {}",
"a rendition of the {}",
"a photo of the clean {}",
"a rendition of a {}",
"a photo of a nice {}",
"a good photo of a {}",
"a photo of the nice {}",
"a photo of the small {}",
"a photo of the weird {}",
"a photo of the large {}",
"a photo of a cool {}",
"a photo of a small {}",
]
imagenet_style_templates_small = [
"a painting in the style of {}",
"a rendering in the style of {}",
"a cropped painting in the style of {}",
"the painting in the style of {}",
"a clean painting in the style of {}",
"a dirty painting in the style of {}",
"a dark painting in the style of {}",
"a picture in the style of {}",
"a cool painting in the style of {}",
"a close-up painting in the style of {}",
"a bright painting in the style of {}",
"a cropped painting in the style of {}",
"a good painting in the style of {}",
"a close-up painting in the style of {}",
"a rendition in the style of {}",
"a nice painting in the style of {}",
"a small painting in the style of {}",
"a weird painting in the style of {}",
"a large painting in the style of {}",
]
class TextualInversionDataset(Dataset):
def __init__(
self,
data_root,
tokenizer,
learnable_property="object", # [object, style]
size=512,
repeats=100,
interpolation="bicubic",
flip_p=0.5,
set="train",
placeholder_token="*",
center_crop=False,
):
self.data_root = Path(data_root)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
self.tokenizer = tokenizer
self.learnable_property = learnable_property
self.size = size
self.placeholder_token = placeholder_token
self.center_crop = center_crop
self.flip_p = flip_p
2023-01-26 20:10:16 +00:00
self.image_paths = [
self.data_root / file_path
for file_path in self.data_root.iterdir()
2023-03-03 06:02:00 +00:00
if file_path.is_file()
and file_path.name.endswith((".png", ".PNG", ".jpg", ".JPG", ".jpeg", ".JPEG", ".gif", ".GIF"))
2023-01-26 20:10:16 +00:00
]
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
self.num_images = len(self.image_paths)
self._length = self.num_images
if set == "train":
self._length = self.num_images * repeats
self.interpolation = {
"linear": PIL_INTERPOLATION["linear"],
"bilinear": PIL_INTERPOLATION["bilinear"],
"bicubic": PIL_INTERPOLATION["bicubic"],
"lanczos": PIL_INTERPOLATION["lanczos"],
}[interpolation]
2023-01-26 20:10:16 +00:00
self.templates = imagenet_style_templates_small if learnable_property == "style" else imagenet_templates_small
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
self.flip_transform = transforms.RandomHorizontalFlip(p=self.flip_p)
def __len__(self):
return self._length
def __getitem__(self, i):
example = {}
image = Image.open(self.image_paths[i % self.num_images])
if not image.mode == "RGB":
image = image.convert("RGB")
placeholder_string = self.placeholder_token
text = random.choice(self.templates).format(placeholder_string)
example["input_ids"] = self.tokenizer(
text,
padding="max_length",
truncation=True,
max_length=self.tokenizer.model_max_length,
return_tensors="pt",
).input_ids[0]
# default to score-sde preprocessing
img = np.array(image).astype(np.uint8)
if self.center_crop:
crop = min(img.shape[0], img.shape[1])
2023-09-01 22:51:53 +00:00
(
h,
w,
) = (
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
img.shape[0],
img.shape[1],
)
2023-01-26 20:10:16 +00:00
img = img[(h - crop) // 2 : (h + crop) // 2, (w - crop) // 2 : (w + crop) // 2]
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
image = Image.fromarray(img)
image = image.resize((self.size, self.size), resample=self.interpolation)
image = self.flip_transform(image)
image = np.array(image).astype(np.uint8)
image = (image / 127.5 - 1.0).astype(np.float32)
example["pixel_values"] = torch.from_numpy(image).permute(2, 0, 1)
return example
2023-01-26 20:10:16 +00:00
def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
if token is None:
token = HfFolder.get_token()
if organization is None:
username = whoami(token)["name"]
return f"{username}/{model_id}"
else:
return f"{organization}/{model_id}"
def do_textual_inversion_training(
config: InvokeAIAppConfig,
2023-01-26 20:10:16 +00:00
model: str,
train_data_dir: Path,
output_dir: Path,
placeholder_token: str,
initializer_token: str,
save_steps: int = 500,
only_save_embeds: bool = False,
revision: str = None,
tokenizer_name: str = None,
learnable_property: str = "object",
repeats: int = 100,
seed: int = None,
resolution: int = 512,
center_crop: bool = False,
train_batch_size: int = 16,
num_train_epochs: int = 100,
max_train_steps: int = 5000,
gradient_accumulation_steps: int = 1,
gradient_checkpointing: bool = False,
learning_rate: float = 1e-4,
scale_lr: bool = True,
lr_scheduler: str = "constant",
lr_warmup_steps: int = 500,
adam_beta1: float = 0.9,
adam_beta2: float = 0.999,
adam_weight_decay: float = 1e-02,
adam_epsilon: float = 1e-08,
push_to_hub: bool = False,
hub_token: str = None,
logging_dir: Path = Path("logs"),
mixed_precision: str = "fp16",
allow_tf32: bool = False,
report_to: str = "tensorboard",
local_rank: int = -1,
checkpointing_steps: int = 500,
resume_from_checkpoint: Path = None,
enable_xformers_memory_efficient_attention: bool = False,
hub_model_id: str = None,
**kwargs,
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
):
2023-01-26 20:10:16 +00:00
assert model, "Please specify a base model with --model"
assert train_data_dir, "Please specify a directory containing the training images using --train_data_dir"
assert placeholder_token, "Please specify a trigger term using --placeholder_token"
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != local_rank:
local_rank = env_local_rank
# setting up things the way invokeai expects them
if not os.path.isabs(output_dir):
output_dir = os.path.join(config.root, output_dir)
2023-01-26 20:10:16 +00:00
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
logging_dir = output_dir / logging_dir
accelerator_config = ProjectConfiguration()
accelerator_config.logging_dir = logging_dir
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
accelerator = Accelerator(
gradient_accumulation_steps=gradient_accumulation_steps,
mixed_precision=mixed_precision,
log_with=report_to,
project_config=accelerator_config,
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
model_manager = ModelManagerService(config, logger)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_warning()
diffusers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
diffusers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if seed is not None:
set_seed(seed)
# Handle the repository creation
if accelerator.is_main_process:
if push_to_hub:
if hub_model_id is None:
repo_name = get_full_repo_name(Path(output_dir).name, token=hub_token)
else:
repo_name = hub_model_id
repo = Repository(output_dir, clone_from=repo_name)
with open(os.path.join(output_dir, ".gitignore"), "w+") as gitignore:
if "step_*" not in gitignore:
gitignore.write("step_*\n")
if "epoch_*" not in gitignore:
gitignore.write("epoch_*\n")
elif output_dir is not None:
os.makedirs(output_dir, exist_ok=True)
known_models = model_manager.model_names()
model_name = model.split("/")[-1]
model_meta = next((mm for mm in known_models if mm[0].endswith(model_name)), None)
assert model_meta is not None, f"Unknown model: {model}"
model_info = model_manager.model_info(*model_meta)
2023-01-26 20:10:16 +00:00
assert model_info["model_format"] == "diffusers", "This script only works with models of type 'diffusers'"
tokenizer_info = model_manager.get_model(*model_meta, submodel=SubModelType.Tokenizer)
noise_scheduler_info = model_manager.get_model(*model_meta, submodel=SubModelType.Scheduler)
text_encoder_info = model_manager.get_model(*model_meta, submodel=SubModelType.TextEncoder)
vae_info = model_manager.get_model(*model_meta, submodel=SubModelType.Vae)
unet_info = model_manager.get_model(*model_meta, submodel=SubModelType.UNet)
pipeline_args = {"local_files_only": True}
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
if tokenizer_name:
2023-01-26 20:10:16 +00:00
tokenizer = CLIPTokenizer.from_pretrained(tokenizer_name, **pipeline_args)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
else:
tokenizer = CLIPTokenizer.from_pretrained(tokenizer_info.location, subfolder="tokenizer", **pipeline_args)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Load scheduler and models
2023-01-26 20:10:16 +00:00
noise_scheduler = DDPMScheduler.from_pretrained(
noise_scheduler_info.location, subfolder="scheduler", **pipeline_args
2023-01-26 20:10:16 +00:00
)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
text_encoder = CLIPTextModel.from_pretrained(
text_encoder_info.location,
2023-01-26 20:10:16 +00:00
subfolder="text_encoder",
revision=revision,
**pipeline_args,
)
vae = AutoencoderKL.from_pretrained(
vae_info.location,
2023-01-26 20:10:16 +00:00
subfolder="vae",
revision=revision,
**pipeline_args,
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
unet = UNet2DConditionModel.from_pretrained(
unet_info.location,
2023-01-26 20:10:16 +00:00
subfolder="unet",
revision=revision,
**pipeline_args,
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
)
# Add the placeholder token in tokenizer
num_added_tokens = tokenizer.add_tokens(placeholder_token)
if num_added_tokens == 0:
raise ValueError(
f"The tokenizer already contains the token {placeholder_token}. Please pass a different"
" `placeholder_token` that is not already in the tokenizer."
)
# Convert the initializer_token, placeholder_token to ids
token_ids = tokenizer.encode(initializer_token, add_special_tokens=False)
# Check if initializer_token is a single token or a sequence of tokens
if len(token_ids) > 1:
2023-01-26 20:10:16 +00:00
raise ValueError(
f"The initializer token must be a single token. Provided initializer={initializer_token}. Token ids={token_ids}"
)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
initializer_token_id = token_ids[0]
placeholder_token_id = tokenizer.convert_tokens_to_ids(placeholder_token)
# Resize the token embeddings as we are adding new special tokens to the tokenizer
text_encoder.resize_token_embeddings(len(tokenizer))
# Initialise the newly added placeholder token with the embeddings of the initializer token
token_embeds = text_encoder.get_input_embeddings().weight.data
token_embeds[placeholder_token_id] = token_embeds[initializer_token_id]
# Freeze vae and unet
vae.requires_grad_(False)
unet.requires_grad_(False)
# Freeze all parameters except for the token embeddings in text encoder
text_encoder.text_model.encoder.requires_grad_(False)
text_encoder.text_model.final_layer_norm.requires_grad_(False)
text_encoder.text_model.embeddings.position_embedding.requires_grad_(False)
if gradient_checkpointing:
# Keep unet in train mode if we are using gradient checkpointing to save memory.
# The dropout cannot be != 0 so it doesn't matter if we are in eval or train mode.
unet.train()
text_encoder.gradient_checkpointing_enable()
unet.enable_gradient_checkpointing()
if enable_xformers_memory_efficient_attention:
if is_xformers_available():
unet.enable_xformers_memory_efficient_attention()
else:
2023-01-26 20:10:16 +00:00
raise ValueError("xformers is not available. Make sure it is installed correctly")
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Enable TF32 for faster training on Ampere GPUs,
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
if allow_tf32:
torch.backends.cuda.matmul.allow_tf32 = True
if scale_lr:
2023-01-26 20:10:16 +00:00
learning_rate = learning_rate * gradient_accumulation_steps * train_batch_size * accelerator.num_processes
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Initialize the optimizer
optimizer = torch.optim.AdamW(
text_encoder.get_input_embeddings().parameters(), # only optimize the embeddings
lr=learning_rate,
betas=(adam_beta1, adam_beta2),
weight_decay=adam_weight_decay,
eps=adam_epsilon,
)
# Dataset and DataLoaders creation:
train_dataset = TextualInversionDataset(
data_root=train_data_dir,
tokenizer=tokenizer,
size=resolution,
placeholder_token=placeholder_token,
repeats=repeats,
learnable_property=learnable_property,
center_crop=center_crop,
set="train",
)
2023-01-26 20:10:16 +00:00
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=train_batch_size, shuffle=True)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
2023-01-26 20:10:16 +00:00
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / gradient_accumulation_steps)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
if max_train_steps is None:
max_train_steps = num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
lr_scheduler,
optimizer=optimizer,
num_warmup_steps=lr_warmup_steps * gradient_accumulation_steps,
num_training_steps=max_train_steps * gradient_accumulation_steps,
)
# Prepare everything with our `accelerator`.
text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
text_encoder, optimizer, train_dataloader, lr_scheduler
)
# For mixed precision training we cast the unet and vae weights to half-precision
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# as these models are only used for inference, keeping weights in full precision is not required.
weight_dtype = torch.float32
if accelerator.mixed_precision == "fp16":
weight_dtype = torch.float16
elif accelerator.mixed_precision == "bf16":
weight_dtype = torch.bfloat16
# Move vae and unet to device and cast to weight_dtype
unet.to(accelerator.device, dtype=weight_dtype)
vae.to(accelerator.device, dtype=weight_dtype)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
2023-01-26 20:10:16 +00:00
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / gradient_accumulation_steps)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
if overrode_max_train_steps:
max_train_steps = num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
num_train_epochs = math.ceil(max_train_steps / num_update_steps_per_epoch)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initializes automatically on the main process.
if accelerator.is_main_process:
params = locals()
2023-01-26 20:10:16 +00:00
for k in params: # init_trackers() doesn't like objects
params[k] = str(params[k]) if isinstance(params[k], object) else params[k]
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
accelerator.init_trackers("textual_inversion", config=params)
# Train!
2023-01-26 20:10:16 +00:00
total_batch_size = train_batch_size * accelerator.num_processes * gradient_accumulation_steps
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {train_batch_size}")
2023-01-26 20:10:16 +00:00
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
logger.info(f" Gradient Accumulation steps = {gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {max_train_steps}")
global_step = 0
first_epoch = 0
resume_step = None
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Potentially load in the weights and states from a previous save
if resume_from_checkpoint:
if resume_from_checkpoint != "latest":
path = os.path.basename(resume_from_checkpoint)
else:
# Get the most recent checkpoint
dirs = os.listdir(output_dir)
dirs = [d for d in dirs if d.startswith("checkpoint")]
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
path = dirs[-1] if len(dirs) > 0 else None
2023-01-26 20:10:16 +00:00
if path is None:
accelerator.print(f"Checkpoint '{resume_from_checkpoint}' does not exist. Starting a new training run.")
resume_from_checkpoint = None
else:
accelerator.print(f"Resuming from checkpoint {path}")
accelerator.load_state(os.path.join(output_dir, path))
global_step = int(path.split("-")[1])
resume_global_step = global_step * gradient_accumulation_steps
first_epoch = global_step // num_update_steps_per_epoch
2023-01-26 20:10:16 +00:00
resume_step = resume_global_step % (num_update_steps_per_epoch * gradient_accumulation_steps)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Only show the progress bar once on each machine.
2023-01-26 20:10:16 +00:00
progress_bar = tqdm(
range(global_step, max_train_steps),
disable=not accelerator.is_local_main_process,
)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
progress_bar.set_description("Steps")
# keep original embeddings as reference
2023-01-26 20:10:16 +00:00
orig_embeds_params = accelerator.unwrap_model(text_encoder).get_input_embeddings().weight.data.clone()
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
for epoch in range(first_epoch, num_train_epochs):
text_encoder.train()
for step, batch in enumerate(train_dataloader):
# Skip steps until we reach the resumed step
2023-01-26 20:10:16 +00:00
if resume_step and resume_from_checkpoint and epoch == first_epoch and step < resume_step:
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
if step % gradient_accumulation_steps == 0:
progress_bar.update(1)
continue
with accelerator.accumulate(text_encoder):
# Convert images to latent space
2023-01-26 20:10:16 +00:00
latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample().detach()
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
latents = latents * 0.18215
# Sample noise that we'll add to the latents
noise = torch.randn_like(latents)
bsz = latents.shape[0]
# Sample a random timestep for each image
2023-01-26 20:10:16 +00:00
timesteps = torch.randint(
0,
noise_scheduler.config.num_train_timesteps,
(bsz,),
device=latents.device,
)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
timesteps = timesteps.long()
# Add noise to the latents according to the noise magnitude at each timestep
# (this is the forward diffusion process)
noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
# Get the text embedding for conditioning
2023-01-26 20:10:16 +00:00
encoder_hidden_states = text_encoder(batch["input_ids"])[0].to(dtype=weight_dtype)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Predict the noise residual
2023-01-26 20:10:16 +00:00
model_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Get the target for loss depending on the prediction type
if noise_scheduler.config.prediction_type == "epsilon":
target = noise
elif noise_scheduler.config.prediction_type == "v_prediction":
target = noise_scheduler.get_velocity(latents, noise, timesteps)
else:
2023-01-26 20:10:16 +00:00
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
accelerator.backward(loss)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
# Let's make sure we don't update any embedding weights besides the newly added token
index_no_updates = torch.arange(len(tokenizer)) != placeholder_token_id
with torch.no_grad():
2023-01-26 20:10:16 +00:00
accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[
index_no_updates
] = orig_embeds_params[index_no_updates]
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
progress_bar.update(1)
global_step += 1
if global_step % save_steps == 0:
2023-01-26 20:10:16 +00:00
save_path = os.path.join(output_dir, f"learned_embeds-steps-{global_step}.bin")
save_progress(
text_encoder,
placeholder_token_id,
accelerator,
placeholder_token,
save_path,
)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
if global_step % checkpointing_steps == 0:
if accelerator.is_main_process:
2023-01-26 20:10:16 +00:00
save_path = os.path.join(output_dir, f"checkpoint-{global_step}")
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
accelerator.save_state(save_path)
logger.info(f"Saved state to {save_path}")
logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
progress_bar.set_postfix(**logs)
accelerator.log(logs, step=global_step)
if global_step >= max_train_steps:
break
# Create the pipeline using using the trained modules and save it.
accelerator.wait_for_everyone()
if accelerator.is_main_process:
if push_to_hub and only_save_embeds:
2023-01-26 20:10:16 +00:00
logger.warn("Enabling full model saving because --push_to_hub=True was specified.")
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
save_full_model = True
else:
save_full_model = not only_save_embeds
if save_full_model:
pipeline = StableDiffusionPipeline.from_pretrained(
unet_info.location,
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
text_encoder=accelerator.unwrap_model(text_encoder),
vae=vae,
unet=unet,
tokenizer=tokenizer,
**pipeline_args,
)
pipeline.save_pretrained(output_dir)
# Save the newly trained embeddings
save_path = os.path.join(output_dir, "learned_embeds.bin")
2023-01-26 20:10:16 +00:00
save_progress(
text_encoder,
placeholder_token_id,
accelerator,
placeholder_token,
save_path,
)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
if push_to_hub:
2023-01-26 20:10:16 +00:00
repo.push_to_hub(commit_message="End of training", blocking=False, auto_lfs_prune=True)
use 🧨diffusers model (#1583) * initial commit of DiffusionPipeline class * spike: proof of concept using diffusers for txt2img * doc: type hints for Generator * refactor(model_cache): factor out load_ckpt * model_cache: add ability to load a diffusers model pipeline and update associated things in Generate & Generator to not instantly fail when that happens * model_cache: fix model default image dimensions * txt2img: support switching diffusers schedulers * diffusers: let the scheduler do its scaling of the initial latents Remove IPNDM scheduler; it is not behaving. * web server: update image_progress callback for diffusers data * diffusers: restore prompt weighting feature * diffusers: fix set-sampler error following model switch * diffusers: use InvokeAIDiffuserComponent for conditioning * cross_attention_control: stub (no-op) implementations for diffusers * model_cache: let offload_model work with DiffusionPipeline, sorta. * models.yaml.example: add diffusers-format model, set as default * test-invoke-conda: use diffusers-format model test-invoke-conda: put huggingface-token where the library can use it * environment-mac: upgrade to diffusers 0.7 (from 0.6) this was already done for linux; mac must have been lost in the merge. * preload_models: explicitly load diffusers models In non-interactive mode too, as long as you're logged in. * fix(model_cache): don't check `model.config` in diffusers format clean-up from recent merge. * diffusers integration: support img2img * dev: upgrade to diffusers 0.8 (from 0.7.1) We get to remove some code by using methods that were factored out in the base class. * refactor: remove backported img2img.get_timesteps now that we can use it directly from diffusers 0.8.1 * ci: use diffusers model * dev: upgrade to diffusers 0.9 (from 0.8.1) * lint: correct annotations for Python 3.9. * lint: correct AttributeError.name reference for Python 3.9. * CI: prefer diffusers-1.4 because it no longer requires a token The RunwayML models still do. * build: there's yet another place to update requirements? * configure: try to download models even without token Models in the CompVis and stabilityai repos no longer require them. (But runwayml still does.) * configure: add troubleshooting info for config-not-found * fix(configure): prepend root to config path * fix(configure): remove second `default: true` from models example * CI: simplify test-on-push logic now that we don't need secrets The "test on push but only in forks" logic was only necessary when tests didn't work for PRs-from-forks. * create an embedding_manager for diffusers * internal: avoid importing diffusers DummyObject see https://github.com/huggingface/diffusers/issues/1479 * fix "config attributes…not expected" diffusers warnings. * fix deprecated scheduler construction * work around an apparent MPS torch bug that causes conditioning to have no effect * 🚧 post-rebase repair * preliminary support for outpainting (no masking yet) * monkey-patch diffusers.attention and use Invoke lowvram code * add always_use_cpu arg to bypass MPS * add cross-attention control support to diffusers (fails on MPS) For unknown reasons MPS produces garbage output with .swap(). Use --always_use_cpu arg to invoke.py for now to test this code on MPS. * diffusers support for the inpainting model * fix debug_image to not crash with non-RGB images. * inpainting for the normal model [WIP] This seems to be performing well until the LAST STEP, at which point it dissolves to confetti. * fix off-by-one bug in cross-attention-control (#1774) prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness). based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly. * refactor common CrossAttention stuff into a mixin so that the old ldm code can still work if necessary * inpainting for the normal model. I think it works this time. * diffusers: reset num_vectors_per_token sync with 44a00555718f1df173c60da0ed646cf700e29537 * diffusers: txt2img2img (hires_fix) with so much slicing and dicing of pipeline methods to stitch them together * refactor(diffusers): reduce some code duplication amongst the different tasks * fixup! refactor(diffusers): reduce some code duplication amongst the different tasks * diffusers: enable DPMSolver++ scheduler * diffusers: upgrade to diffusers 0.10, add Heun scheduler * diffusers(ModelCache): stopgap to make from_cpu compatible with diffusers * CI: default to diffusers-1.5 now that runwayml token requirement is gone * diffusers: update to 0.10 (and transformers to 4.25) * diffusers: use xformers when available diffusers no longer auto-enables this as of 0.10.2. * diffusers: make masked img2img behave better with multi-step schedulers re-randomizing the noise each step was confusing them. * diffusers: work more better with more models. fixed relative path problem with local models. fixed models on hub not always having a `fp16` branch. * diffusers: stopgap fix for attention_maps_callback crash after recent merge * fixup import merge conflicts correction for 061c5369a2247c6c92cd69606bcf54c4f1962a0b * test: add tests/inpainting inputs for masked img2img * diffusers(AddsMaskedGuidance): partial fix for k-schedulers Prevents them from crashing, but results are still hot garbage. * fix --safety_checker arg parsing and add note to diffusers loader about where safety checker gets called * generate: fix import error * CI: don't try to read the old init location * diffusers: support loading an alternate VAE * CI: remove sh-syntax if-statement so it doesn't crash powershell * CI: fold strings in yaml because backslash is not line-continuation in powershell * attention maps callback stuff for diffusers * build: fix syntax error in environment-mac * diffusers: add INITIAL_MODELS with diffusers-compatible repos * re-enable the embedding manager; closes #1778 * Squashed commit of the following: commit e4a956abc37fcb5cf188388b76b617bc5c8fda7d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:43:07 2022 +0100 import new load handling from EmbeddingManager and cleanup commit c4abe91a5ba0d415b45bf734068385668b7a66e6 Merge: 032e856e 1efc6397 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:09:53 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers_with_textual_inversion_manager commit 032e856eefb3bbc39534f5daafd25764bcfcef8b Merge: 8b4f0fe9 bc515e24 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:08:01 2022 +0100 Merge remote-tracking branch 'upstream/dev/diffusers' into dev/diffusers_with_textual_inversion_manager commit 1efc6397fc6e61c1aff4b0258b93089d61de5955 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 15:04:28 2022 +0100 cleanup and add performance notes commit e400f804ac471a0ca2ba432fd658778b20c7bdab Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:45:07 2022 +0100 fix bug and update unit tests commit deb9ae0ae1016750e93ce8275734061f7285a231 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 14:28:29 2022 +0100 textual inversion manager seems to work commit 162e02505dec777e91a983c4d0fb52e950d25ff0 Merge: cbad4583 12769b3d Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:58:03 2022 +0100 Merge branch 'main' into feature_textual_inversion_mgr commit cbad45836c6aace6871a90f2621a953f49433131 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:54:10 2022 +0100 use position embeddings commit 070344c69b0e0db340a183857d0a787b348681d3 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:53:47 2022 +0100 Don't crash CLI on exceptions commit b035ac8c6772dfd9ba41b8eeb9103181cda028f8 Author: Damian Stewart <d@damianstewart.com> Date: Sun Dec 18 11:11:55 2022 +0100 add missing position_embeddings commit 12769b3d3562ef71e0f54946b532ad077e10043c Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:33:25 2022 +0100 debugging why it don't work commit bafb7215eabe1515ca5e8388fd3bb2f3ac5362cf Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 13:21:33 2022 +0100 debugging why it don't work commit 664a6e9e146b42d96703f0cc8baf8f5efec04ee1 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit 8b4f0fe9d6e4e2643b36dfa27864294785d7ba4e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 12:48:38 2022 +0100 use TextualInversionManager in place of embeddings (wip, doesn't work) commit ffbe1ab11163ba712e353d89404e301d0e0c6cdf Merge: 6e4dad60 023df37e Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:37:31 2022 +0100 Merge branch 'feature_textual_inversion_mgr' into dev/diffusers commit 023df37efffa67434f77def7fc3c9dfb29f699fd Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:36:54 2022 +0100 cleanup commit 05fac594eaf79d0058e3c48deee93df603f136c2 Author: Damian Stewart <d@damianstewart.com> Date: Fri Dec 16 02:07:49 2022 +0100 tweak error checking commit 009f32ed39a7280997c3ffab112adadee0b44279 Author: damian <null@damianstewart.com> Date: Thu Dec 15 21:29:47 2022 +0100 unit tests passing for embeddings with vector length >1 commit beb1b08d9a98112ed2fe073580568e1a18698da3 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:39:09 2022 +0100 more explicit equality tests when overwriting commit 44d8a5a7c85cdabc9ce3a54fd0769a10597b3ca9 Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 13:30:13 2022 +0100 wip textual inversion manager (unit tests passing for 1v embedding overwriting) commit 417c2b57d90924a839616bfb66804faab8039e4c Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 12:30:55 2022 +0100 wip textual inversion manager (unit tests passing for base stuff + padding) commit 2e80872e3b6f7fd7d8eb8928822bd824b63cb2ff Author: Damian Stewart <d@damianstewart.com> Date: Thu Dec 15 10:57:57 2022 +0100 wip new TextualInversionManager * stop using WeightedFrozenCLIPEmbedder * store diffusion models locally - configure_invokeai.py reconfigured to store diffusion models rather than CompVis models - hugging face caching model is used, but cache is set to ~/invokeai/models/repo_id - models.yaml does **NOT** use path, just repo_id - "repo_name" changed to "repo_id" to following hugging face conventions - Models are loaded with full precision pending further work. * allow non-local files during development * path takes priority over repo_id * MVP for model_cache and configure_invokeai - Feature complete (almost) - configure_invokeai.py downloads both .ckpt and diffuser models, along with their VAEs. Both types of download are controlled by a unified INITIAL_MODELS.yaml file. - model_cache can load both type of model and switches back and forth in CPU. No memory leaks detected TO DO: 1. I have not yet turned on the LocalOnly flag for diffuser models, so the code will check the Hugging Face repo for updates before using the locally cached models. This will break firewalled systems. I am thinking of putting in a global check for internet connectivity at startup time and setting the LocalOnly flag based on this. It would be good to check updates if there is connectivity. 2. I have not gone completely through INITIAL_MODELS.yaml to check which models are available as diffusers and which are not. So models like PaperCut and VoxelArt may not load properly. The runway and stability models are checked, as well as the Trinart models. 3. Add stanzas for SD 2.0 and 2.1 in INITIAL_MODELS.yaml REMAINING PROBLEMS NOT DIRECTLY RELATED TO MODEL_CACHE: 1. When loading a .ckpt file there are lots of messages like this: Warning! ldm.modules.attention.CrossAttention is no longer being maintained. Please use InvokeAICrossAttention instead. I'm not sure how to address this. 2. The ckpt models ***don't actually run*** due to the lack of special-case support for them in the generator objects. For example, here's the hard crash you get when you run txt2img against the legacy waifu-diffusion-1.3 model: ``` >> An error occurred: Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 140, in main main_loop(gen, opt) File "/data/lstein/InvokeAI/ldm/invoke/CLI.py", line 371, in main_loop gen.prompt2image( File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1265, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'LatentDiffusion' object has no attribute 'image_from_embeddings' ``` 3. The inpainting diffusion model isn't working. Here's the output of "banana sushi" when inpainting-1.5 is loaded: ``` Traceback (most recent call last): File "/data/lstein/InvokeAI/ldm/generate.py", line 496, in prompt2image results = generator.generate( File "/data/lstein/InvokeAI/ldm/invoke/generator/base.py", line 108, in generate image = make_image(x_T) File "/data/lstein/InvokeAI/ldm/invoke/generator/txt2img.py", line 33, in make_image pipeline_output = pipeline.image_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 301, in image_from_embeddings result_latents, result_attention_map_saver = self.latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 330, in latents_from_embeddings result: PipelineIntermediateState = infer_latents_from_embeddings( File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 185, in __call__ for result in self.generator_method(*args, **kwargs): File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 367, in generate_latents_from_embeddings step_output = self.step(batched_t, latents, guidance_scale, File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/data/lstein/InvokeAI/ldm/invoke/generator/diffusers_pipeline.py", line 409, in step step_output = self.scheduler.step(noise_pred, timestep, latents, **extra_step_kwargs) File "/home/lstein/invokeai/.venv/lib/python3.9/site-packages/diffusers/schedulers/scheduling_lms_discrete.py", line 223, in step pred_original_sample = sample - sigma * model_output RuntimeError: The size of tensor a (9) must match the size of tensor b (4) at non-singleton dimension 1 ``` * proper support for float32/float16 - configure script now correctly detects user's preference for fp16/32 and downloads the correct diffuser version. If fp16 version not available, falls back to fp32 version. - misc code cleanup and simplification in model_cache * add on-the-fly conversion of .ckpt to diffusers models 1. On-the-fly conversion code can be found in the file ldm/invoke/ckpt_to_diffusers.py. 2. A new !optimize command has been added to the CLI. Should be ported to Web GUI. User experience on the CLI is this: ``` invoke> !optimize /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt INFO: Converting legacy weights file /home/lstein/invokeai/models/ldm/stable-diffusion-v1/sd-v1-4.ckpt to optimized diffuser model. This operation will take 30-60s to complete. Success. Optimized model is now located at /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 Writing new config file entry for sd-v1-4... >> New configuration: sd-v1-4: description: Optimized version of sd-v1-4 format: diffusers path: /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 OK to import [n]? y >> Verifying that new model loads... >> Current VRAM usage: 2.60G >> Offloading stable-diffusion-2.1 to CPU >> Loading diffusers model from /home/lstein/tmp/invokeai/models/optimized-ckpts/sd-v1-4 | Using faster float16 precision You have disabled the safety checker for <class 'ldm.invoke.generator.diffusers_pipeline.StableDiffusionGeneratorPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion \ license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances,\ disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 . | training width x height = (512 x 512) >> Model loaded in 3.48s >> Max VRAM used to load the model: 2.17G >> Current VRAM usage:2.17G >> Textual inversions available: >> Setting Sampler to k_lms (LMSDiscreteScheduler) Keep model loaded? [y] ``` * add parallel set of generator files for ckpt legacy generation * generation using legacy ckpt models now working * diffusers: fix missing attention_maps_callback fix for 23eb80b40421b2bb8f4b6d3dd30490d11c447b36 * associate legacy CrossAttention with .ckpt models * enable autoconvert New --autoconvert CLI option will scan a designated directory for new .ckpt files, convert them into diffuser models, and import them into models.yaml. Works like this: invoke.py --autoconvert /path/to/weights/directory In ModelCache added two new methods: autoconvert_weights(config_path, weights_directory_path, models_directory_path) convert_and_import(ckpt_path, diffuser_path) * diffusers: update to diffusers 0.11 (from 0.10.2) * fix vae loading & width/height calculation * refactor: encapsulate these conditioning data into one container * diffusers: fix some noise-scaling issues by pushing the noise-mixing down to the common function * add support for safetensors and accelerate * set local_files_only when internet unreachable * diffusers: fix error-handling path when model repo has no fp16 branch * fix generatorinpaint error Fixes : "ModuleNotFoundError: No module named 'ldm.invoke.generatorinpaint' https://github.com/invoke-ai/InvokeAI/pull/1583#issuecomment-1363634318 * quench diffuser safety-checker warning * diffusers: support stochastic DDIM eta parameter * fix conda env creation on macos * fix cross-attention with diffusers 0.11 * diffusers: the VAE needs to be tiling as well as the U-Net * diffusers: comment on subfolders * diffusers: embiggen! * diffusers: make model_cache.list_models serializable * diffusers(inpaint): restore scaling functionality * fix requirements clash between numba and numpy 1.24 * diffusers: allow inpainting model to do non-inpainting tasks * start expanding model_cache functionality * add import_ckpt_model() and import_diffuser_model() methods to model_manager - in addition, model_cache.py is now renamed to model_manager.py * allow "recommended" flag to be optional in INITIAL_MODELS.yaml * configure_invokeai now downloads VAE diffusers in advance * rename ModelCache to ModelManager * remove support for `repo_name` in models.yaml * check for and refuse to load embeddings trained on incompatible models * models.yaml.example: s/repo_name/repo_id and remove extra INITIAL_MODELS now that the main one has diffusers models in it. * add MVP textual inversion script * refactor(InvokeAIDiffuserComponent): factor out _combine() * InvokeAIDiffuserComponent: implement threshold * InvokeAIDiffuserComponent: diagnostic logs for threshold ...this does not look right * add a curses-based frontend to textual inversion - not quite working yet - requires npyscreen installed - on windows will also have the windows-curses requirement, but not added to requirements yet * add curses-based interface for textual inversion * fix crash in convert_and_import() - This corrects a "local variable referenced before assignment" error in model_manager.convert_and_import() * potential workaround for no 'state_dict' key error - As reported in https://github.com/huggingface/diffusers/issues/1876 * create TI output dir if needed * Update environment-lin-cuda.yml (#2159) Fixing line 42 to be the proper order to define the transformers requirement: ~= instead of =~ * diffusers: update sampler-to-scheduler mapping based on https://github.com/huggingface/diffusers/issues/277#issuecomment-1371428672 * improve user exp for ckt to diffusers conversion - !optimize_models command now operates on an existing ckpt file entry in models.yaml - replaces existing entry, rather than adding a new one - offers to delete the ckpt file after conversion * web: adapt progress callback to deal with old generator or new diffusers pipeline * clean-up model_manager code - add_model() verified to work for .ckpt local paths, .ckpt remote URLs, diffusers local paths, and diffusers repo_ids - convert_and_import() verified to work for local and remove .ckpt files * handle edge cases for import_model() and convert_model() * add support for safetensor .ckpt files * fix name error * code cleanup with pyflake * improve model setting behavior - If the user enters an invalid model name at startup time, will not try to load it, warn, and use default model - CLI UI enhancement: include currently active model in the command line prompt. * update test-invoke-pip.yml - fix model cache path to point to runwayml/stable-diffusion-v1-5 - remove `skip-sd-weights` from configure_invokeai.py args * exclude dev/diffusers from "fail for draft PRs" * disable "fail on PR jobs" * re-add `--skip-sd-weights` since no space * update workflow environments - include `INVOKE_MODEL_RECONFIGURE: '--yes'` * clean up model load failure handling - Allow CLI to run even when no model is defined or loadable. - Inhibit stack trace when model load fails - only show last error - Give user *option* to run configure_invokeai.py when no models successfully load. - Restart invokeai after reconfiguration. * further edge-case handling 1) only one model in models.yaml file, and that model is broken 2) no models in models.yaml 3) models.yaml doesn't exist at all * fix incorrect model status listing - "cached" was not being returned from list_models() - normalize handling of exceptions during model loading: - Passing an invalid model name to generate.set_model() will return a KeyError - All other exceptions are returned as the appropriate Exception * CI: do download weights (if not already cached) * diffusers: fix scheduler loading in offline mode * CI: fix model name (no longer has `diffusers-` prefix) * Update txt2img2img.py (#2256) * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * feat - make model storage compatible with hugging face caching system This commit alters the InvokeAI model directory to be compatible with hugging face, making it easier to share diffusers (and other models) across different programs. - If the HF_HOME environment variable is not set, then models are cached in ~/invokeai/models in a format that is identical to the HuggingFace cache. - If HF_HOME is set, then models are cached wherever HF_HOME points. - To enable sharing with other HuggingFace library clients, set HF_HOME to ~/.cache/huggingface to set the default cache location or to ~/invokeai/models to have huggingface cache inside InvokeAI. * fixes to share models with HuggingFace cache system - If HF_HOME environment variable is defined, then all huggingface models are stored in that directory following the standard conventions. - For seamless interoperability, set HF_HOME to ~/.cache/huggingface - If HF_HOME not defined, then models are stored in ~/invokeai/models. This is equivalent to setting HF_HOME to ~/invokeai/models A future commit will add a migration mechanism so that this change doesn't break previous installs. * fix error "no attribute CkptInpaint" * model_manager.list_models() returns entire model config stanza+status * Initial Draft - Model Manager Diffusers * added hash function to diffusers * implement sha256 hashes on diffusers models * Add Model Manager Support for Diffusers * fix various problems with model manager - in cli import functions, fix not enough values to unpack from _get_name_and_desc() - fix crash when using old-style vae: value with new-style diffuser * rebuild frontend * fix dictconfig-not-serializable issue * fix NoneType' object is not subscriptable crash in model_manager * fix "str has no attribute get" error in model_manager list_models() * Add path and repo_id support for Diffusers Model Manager Also fixes bugs * Fix tooltip IT localization not working * Add Version Number To WebUI * Optimize Model Search * Fix incorrect font on the Model Manager UI * Fix image degradation on merge fixes - [Experimental] This change should effectively fix a couple of things. - Fix image degradation on subsequent merges of the canvas layers. - Fix the slight transparent border that is left behind when filling the bounding box with a color. - Fix the left over line of color when filling a bounding box with color. So far there are no side effects for this. If any, please report. * Add local model filtering for Diffusers / Checkpoints * Go to home on modal close for the Add Modal UI * Styling Fixes * Model Manager Diffusers Localization Update * Add Safe Tensor scanning to Model Manager * Fix model edit form dispatching string values instead of numbers. * Resolve VAE handling / edge cases for supplied repos * defer injecting tokens for textual inversions until they're used for the first time * squash a console warning * implement model migration check * add_model() overwrites previous config rather than merges * fix model config file attribute merging * fix precision handling in textual inversion script * allow ckpt conversion script to work with safetensors .ckpts Applied patch here: https://github.com/huggingface/diffusers/commit/beb932c5d111872c5e45387e7b1b2b3dd0524a47 * fix name "args" is not defined crash in textual_inversion_training * fix a second NameError: name 'args' is not defined crash * fix loading of the safety checker from the global cache dir * add installation step to textual inversion frontend - After a successful training run, the script will copy learned_embeds.bin to a subfolder of the embeddings directory. - User given the option to delete the logs and intermediate checkpoints (which together use 7-8G of space) - If textual inversion training fails, reports the error gracefully. * don't crash out on incompatible embeddings - put try: blocks around places where the system tries to load an embedding which is incompatible with the currently loaded model * add support for checkpoint resuming * textual inversion preferences are saved and restored between sessions - Preferences are stored in a file named text-inversion-training/preferences.conf - Currently the resume-from-checkpoint option is not working correctly. Possible bug in textual_inversion_training.py? * copy learned_embeddings.bin into right location * add front end for diffusers model merging - Front end doesn't do anything yet!!!! - Made change to model name parsing in CLI to support ability to have merged models with the "+" character in their names. * improve inpainting experience - recommend ckpt version of inpainting-1.5 to user - fix get_noise() bug in ckpt version of omnibus.py * update environment*yml * tweak instructions to install HuggingFace token * bump version number * enhance update scripts - update scripts will now fetch new INITIAL_MODELS.yaml so that configure_invokeai.py will know about the diffusers versions. * enhance invoke.sh/invoke.bat launchers - added configure_invokeai.py to menu - menu defaults to browser-based invoke * remove conda workflow (#2321) * fix `token_ids has shape torch.Size([79]) - expected [77]` * update CHANGELOG.md with 2.3.* info - Add information on how formats have changed and the upgrade process. - Add short bug list. Co-authored-by: Damian Stewart <d@damianstewart.com> Co-authored-by: Damian Stewart <null@damianstewart.com> Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com> Co-authored-by: Wybartel-luxmc <37852506+Wybartel-luxmc@users.noreply.github.com> Co-authored-by: mauwii <Mauwii@outlook.de> Co-authored-by: mickr777 <115216705+mickr777@users.noreply.github.com> Co-authored-by: blessedcoolant <54517381+blessedcoolant@users.noreply.github.com> Co-authored-by: Eugene Brodsky <ebr@users.noreply.github.com> Co-authored-by: Matthias Wild <40327258+mauwii@users.noreply.github.com>
2023-01-15 14:22:46 +00:00
accelerator.end_training()