Mirrored_Repos/InvokeAI
1
0
Fork 0
mirror of https://github.com/invoke-ai/InvokeAI synced 2024-08-30 20:32:17 +00:00
Code Issues Packages Projects Releases Wiki Activity
12,667 Commits 274 Branches 239 Tags 397 MiB
e0f12c762e
Commit Graph

1827 Commits

Author SHA1 Message Date
Sergey Borisov
42356ec866 Add ControlNet support to denoise 2024-07-21 20:01:30 +03:00
Sergey Borisov
1748848b7b Ruff fixes 2024-07-21 18:37:20 +03:00
Sergey Borisov
5772965f09 Fix slightly different output with old backend 2024-07-21 18:31:30 +03:00
Sergey Borisov
e046e60e1c Add FreeU support to denoise 2024-07-21 18:31:10 +03:00
Sergey Borisov
9a1420280e Add rescale cfg support to denoise 2024-07-21 17:33:43 +03:00
Mary Hipp
930ff559e4 add sdxl tile to starter models 2024-07-19 16:49:33 -04:00
Ryan Dick
78d2b1b650 Merge branch 'main' into stalker-backend_base 2024-07-19 16:25:20 -04:00
Sergey Borisov
39e10d894c Add invocation cancellation logic to patchers 2024-07-19 23:17:01 +03:00
Ryan Dick
e16faa6370 Add gradient blending to tile seams in MultiDiffusion. 2024-07-19 13:05:50 -07:00
Ryan Dick
83a86abce2 Add unit tests for ExtensionsManager and ExtensionBase. 2024-07-19 14:15:46 -04:00
Sergey Borisov
0c56d4a581 Ryan's suggested changes to extension manager/extensions 
Co-Authored-By: Ryan Dick <14897797+RyanJDick@users.noreply.github.com>
 2024-07-18 23:49:44 +03:00
Lincoln Stein
97a7f51721
don't use cpu state_dict for model unpatching when executing on cpu (#6631) 
Co-authored-by: Lincoln Stein <lstein@gmail.com>
 2024-07-18 15:34:01 -04:00
StAlKeR7779
710dc6b487
Merge branch 'main' into stalker7779/backend_base 2024-07-18 01:08:04 +03:00
Sergey Borisov
2ef3b49a79 Add run cancelling logic to extension manager 2024-07-17 04:39:15 +03:00
Sergey Borisov
3f79467f7b Ruff format 2024-07-17 04:24:45 +03:00
Sergey Borisov
2c2ec8f0bc Comments, a bit refactor 2024-07-17 04:20:31 +03:00
Sergey Borisov
79e35bd0d3 Minor fixes 2024-07-17 03:48:37 +03:00
Sergey Borisov
137202b77c Remove patch_unet logic for now 2024-07-17 03:40:27 +03:00
Sergey Borisov
03e22c257b Convert conditioning_mode to enum 2024-07-17 03:37:11 +03:00
Sergey Borisov
ae6d4fbc78 Move out _concat_conditionings_for_batch submethods 2024-07-17 03:31:26 +03:00
Sergey Borisov
cd1bc1595a Rename sequential as private variable 2024-07-17 03:24:11 +03:00
Ryan Dick
f866b49255 Add some ESRGAN and SwinIR upscale models to the starter models list. 2024-07-16 15:55:10 -04:00
Sergey Borisov
b7c6c63005 Added some comments 2024-07-16 22:52:44 +03:00
Ryan Dick
6b0ca88177
Merge branch 'main' into ryan/spandrel-upscale-tiling 2024-07-16 15:40:14 -04:00
Ryan Dick
81991e072b Merge branch 'main' into ryan/spandrel-upscale 2024-07-16 15:14:08 -04:00
Sergey Borisov
cec345cb5c Change attention processor apply logic 2024-07-16 20:03:29 +03:00
Sergey Borisov
608cbe3f5c Separate inputs in denoise context 2024-07-16 19:30:29 +03:00
psychedelicious
38343917f8 fix(backend): revert non-blocking device transfer 
In #6490 we enabled non-blocking torch device transfers throughout the model manager's memory management code. When using this torch feature, torch attempts to wait until the tensor transfer has completed before allowing any access to the tensor. Theoretically, that should make this a safe feature to use.

This provides a small performance improvement but causes race conditions in some situations. Specific platforms/systems are affected, and complicated data dependencies can make this unsafe.

- Intermittent black images on MPS devices - reported on discord and #6545, fixed with special handling in #6549.
- Intermittent OOMs and black images on a P4000 GPU on Windows - reported in #6613, fixed in this commit.

On my system, I haven't experience any issues with generation, but targeted testing of non-blocking ops did expose a race condition when moving tensors from CUDA to CPU.

One workaround is to use torch streams with manual sync points. Our application logic is complicated enough that this would be a lot of work and feels ripe for edge cases and missed spots.

Much safer is to fully revert non-locking - which is what this change does.
 2024-07-16 08:59:42 +10:00
Sergey Borisov
9f088d1bf5 Multiple small fixes 2024-07-16 00:51:25 +03:00
Sergey Borisov
fd8d1c12d4 Remove 'del' operator overload 2024-07-16 00:43:32 +03:00
Sergey Borisov
d623bd429b Fix condtionings logic 2024-07-16 00:31:56 +03:00
psychedelicious
28e79c4c5e chore: ruff 
Looks like an upstream change to ruff resulted in this file being a violation.
 2024-07-15 14:05:04 +10:00
Sergey Borisov
499e4d4fde Add preview extension to check logic 2024-07-13 00:45:04 +03:00
Sergey Borisov
e961dd1dec Remove remains of priority logic 2024-07-13 00:44:21 +03:00
Sergey Borisov
7e00526999 Remove overrides logic for now 2024-07-13 00:28:56 +03:00
Sergey Borisov
3a9dda9177 Renames 2024-07-12 22:44:00 +03:00
Sergey Borisov
bd8ae5d896 Simplify guidance modes 2024-07-12 22:01:37 +03:00
Sergey Borisov
87e96e1be2 Rename modifiers to callbacks, convert order to int, a bit unify injection points 2024-07-12 22:01:05 +03:00
Sergey Borisov
0bc60378d3 A bit rework conditioning convert to unet kwargs 2024-07-12 20:43:32 +03:00
Sergey Borisov
9cc852cf7f Base code from draft PR 2024-07-12 20:31:26 +03:00
Ryan Dick
ab775726b7 Add tiling support to the SpoandrelImageToImage node. 2024-07-10 14:25:19 -04:00
Ryan Dick
650902dc29 Fix broken unit test caused by non-existent model path. 2024-07-10 13:59:17 -04:00
Ryan Dick
7b5d4935b4 Merge branch 'main' into ryan/spandrel-upscale 2024-07-09 13:47:11 -04:00
Ryan Dick
af63c538ed Demote error log to warning to models treated as having size 0. 2024-07-09 08:35:43 -04:00
Ryan Dick
0ce6ec634d Do not assign the result of SpandrelImageToImageModel.load_from_file(...) during probe to ensure that the model is immediately gc'd. 2024-07-05 14:05:12 -04:00
Ryan Dick
35f8781ea2 Fix static type errors with SCHEDULER_NAME_VALUES. And, avoid bi-directional cross-directory imports, which contribute to circular import issues. 2024-07-05 07:38:35 -07:00
Ryan Dick
36202d6d25 Delete unused duplicate libc_util.py file. The active version is at invokeai/backend/model_manager/libc_util.py. 2024-07-04 10:30:40 -04:00
Ryan Dick
1d449097cc Apply ruff rule to disallow all relative imports. 2024-07-04 09:35:37 -04:00
Ryan Dick
9da5925287 Add ruff rule to disallow relative parent imports. 2024-07-04 09:35:37 -04:00
Ryan Dick
414750a45d Update calc_model_size_by_data(...) to handle all expected model types, and to log an error if an unexpected model type is received. 2024-07-04 09:08:25 -04:00
Ryan Dick
a405f14ea2 Fix SpandrelImageToImageModel size calculation for the model cache. 2024-07-03 16:38:16 -04:00
Ryan Dick
114320ee69 (minor) typo 2024-07-03 16:28:21 -04:00
Ryan Dick
6161aa73af Move pil_to_tensor() and tensor_to_pil() utilities to the SpandrelImageToImage class. 2024-07-03 16:28:21 -04:00
Ryan Dick
1ab20f43c8 Tidy spandrel model probe logic, and document the reasons behind the current implementation. 2024-07-03 16:28:21 -04:00
Ryan Dick
29c8ddfb88 WIP - A bunch of boilerplate to support Spandrel Image-to-Image models throughout the model manager and the frontend. 2024-07-03 16:28:21 -04:00
Ryan Dick
95079dc7d4 Use a ModelIdentifierField to identify the spandrel model in the UpscaleSpandrelInvocation. 2024-07-03 16:28:21 -04:00
Ryan Dick
2a1514272f Set the dtype correctly for SpandrelImageToImageModels when they are loaded. 2024-07-03 16:28:21 -04:00
Ryan Dick
59ce9cf41c WIP - Begin to integrate SpandreImageToImageModel type into the model manager. 2024-07-03 16:28:21 -04:00
Ryan Dick
e6abea7bc5 (minor) Remove redundant else clause on a for-loop with no break statement. 2024-07-03 16:28:21 -04:00
Ryan Dick
c335f92345 (minor) simplify startswith(...) syntax. 2024-07-03 16:28:21 -04:00
Ryan Dick
e4813f800a Update calc_model_size_by_data(...) to handle all expected model types, and to log an error if an unexpected model type is received. 2024-07-02 21:51:45 -04:00
Ryan Dick
3752509066 Expose the VAE tile_size on the VAE encode and decode invocations. 2024-07-02 09:07:03 -04:00
Ryan Dick
79640ba14e Add context manager for overriding VAE tiling params. 2024-07-02 09:07:03 -04:00
Kent Keirsey
5df2a79549 Update starter models 2024-06-28 17:49:45 +10:00
Kent Keirsey
10b9088312 update controlnet starter models 2024-06-28 17:49:45 +10:00
Lincoln Stein
3e0fb45dd7
Load single-file checkpoints directly without conversion (#6510) 
* use model_class.load_singlefile() instead of converting; works, but performance is poor

* adjust the convert api - not right just yet

* working, needs sql migrator update

* rename migration_11 before conflict merge with main

* Update invokeai/backend/model_manager/load/model_loaders/stable_diffusion.py

Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>

* Update invokeai/backend/model_manager/load/model_loaders/stable_diffusion.py

Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>

* implement lightweight version-by-version config migration

* simplified config schema migration code

* associate sdxl config with sdxl VAEs

* remove use of original_config_file in load_single_file()

---------

Co-authored-by: Lincoln Stein <lstein@gmail.com>
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
 2024-06-27 17:31:28 -04:00
Ryan Dick
14775cc9c4 ruff format 2024-06-27 09:45:13 -04:00
psychedelicious
c7562dd6c0
fix(backend): mps should not use non_blocking 
We can get black outputs when moving tensors from CPU to MPS. It appears MPS to CPU is fine. See:
- https://github.com/pytorch/pytorch/issues/107455
- https://discuss.pytorch.org/t/should-we-set-non-blocking-to-true/38234/28

Changes:
- Add properties for each device on `TorchDevice` as a convenience.
- Add `get_non_blocking` static method on `TorchDevice`. This utility takes a torch device and returns the flag to be used for non_blocking when moving a tensor to the device provided.
- Update model patching and caching APIs to use this new utility.

Fixes: #6545
 2024-06-27 19:15:23 +10:00
Ryan Dick
9a3b8c6fcb Fix handling of init_timestep in StableDiffusionGeneratorPipeline and improve its documentation. 2024-06-26 12:51:51 -04:00
Ryan Dick
bd74b84cc5 Revert "Remove the redundant init_timestep parameter that was being passed around. It is simply the first element of the timesteps array." 
This reverts commit fa40061eca.
 2024-06-26 12:51:51 -04:00
Brandon Rising
dc23bebebf Run ruff 2024-06-26 21:46:59 +10:00
Kent Keirsey
38b6f90c02 Update prevention exception message 2024-06-26 21:46:59 +10:00
Ryan Dick
cd9dfefe3c Fix inpainting mask shape assertions. 2024-06-25 11:31:52 -07:00
Ryan Dick
e1af78c702 Make the tile_overlap input to MultiDiffusion *strictly* control the amount of overlap rather than being a lower bound. 2024-06-25 11:31:52 -07:00
Ryan Dick
c5588e1ff7 Add TODO comment explaining why some schedulers do not interact well with MultiDiffusion. 2024-06-25 11:31:52 -07:00
Ryan Dick
07ac292680 Consolidate _region_step() function - the separation wasn't really adding any value. 2024-06-25 11:31:52 -07:00
Ryan Dick
7c032ea604 (minor) Fix some documentation typos. 2024-06-25 11:31:52 -07:00
Ryan Dick
fa40061eca Remove the redundant init_timestep parameter that was being passed around. It is simply the first element of the timesteps array. 2024-06-25 11:31:52 -07:00
Ryan Dick
25067e4f0d Delete rough notes. 2024-06-25 11:31:52 -07:00
Ryan Dick
fb0aaa3e6d Fix advanced scheduler behaviour in MultiDiffusionPipeline. 2024-06-25 11:31:52 -07:00
Ryan Dick
c22526b9d0 Fix handling of stateful schedulers in MultiDiffusionPipeline. 2024-06-25 11:31:52 -07:00
Ryan Dick
c881882f73 Connect TiledMultiDiffusionDenoiseLatents to the MultiDiffusionPipeline backend. 2024-06-25 11:31:52 -07:00
Ryan Dick
36473fc52a Remove regional conditioning logic from MultiDiffusionPipeline - it is not yet supported. 2024-06-25 11:31:52 -07:00
Ryan Dick
b9964ecc4a Initial (untested) implementation of MultiDiffusionPipeline. 2024-06-25 11:31:52 -07:00
Ryan Dick
051af802fe Remove inpainting support from MultiDiffusionPipeline. 2024-06-25 11:31:52 -07:00
Ryan Dick
3ff2e558d9 Remove IP-Adapter and T2I-Adapter support from MultiDiffusionPipeline. 2024-06-25 11:31:52 -07:00
Ryan Dick
fc187c9253 Document plan for the rest of the MultiDiffusion implementation. 2024-06-25 11:31:52 -07:00
Ryan Dick
605f460c7d Add detailed docstring to latents_from_embeddings(). 2024-06-25 11:31:52 -07:00
Ryan Dick
60d1e686d8 Copy StableDiffusionGeneratorPipeline as a starting point for a new MultiDiffusionPipeline. 2024-06-25 11:31:52 -07:00
Ryan Dick
22704dd542 Simplify handling of inpainting models. Improve the in-code documentation around inpainting. 2024-06-25 11:31:52 -07:00
Ryan Dick
875673c9ba Minor tidying of latents_from_embeddings(...). 2024-06-25 11:31:52 -07:00
Ryan Dick
f604575862 Consolidate latents_from_embeddings(...) and generate_latents_from_embeddings(...) into a single function. 2024-06-25 11:31:52 -07:00
Ryan Dick
60ac937698 Improve clarity of comments regarded when 'noise' and 'latents' are expected to be set. 2024-06-25 11:31:52 -07:00
Ryan Dick
1e41949a02 Fix static check errors on imports in diffusers_pipeline.py. 2024-06-25 11:31:52 -07:00
Ryan Dick
5f0e330ed2 Remove a condition for handling inpainting models that never resolves to True. The same logic is already applied earlier by AddsMaskLatents. 2024-06-25 11:31:52 -07:00
Ryan Dick
9dd779b414 Add clarifying comment to explain why noise might be None in latents_from_embedding(). 2024-06-25 11:31:52 -07:00
Ryan Dick
fa183025ac Remove unused are_like_tensors() function. 2024-06-25 11:31:52 -07:00
Ryan Dick
d3c85aa91a Remove unused StableDiffusionGeneratorPipeline.use_ip_adapter member. 2024-06-25 11:31:52 -07:00
Ryan Dick
82619602a5 Remove unused StableDiffusionGeneratorPipeline.control_model. 2024-06-25 11:31:52 -07:00
Ryan Dick
196f3b721d Stricter typing for the is_gradient_mask: bool. 2024-06-25 11:31:52 -07:00
Ryan Dick
244c28859d Fix typing of control_data to reflect that it can be None. 2024-06-25 11:31:52 -07:00
Ryan Dick
40ae174c41 Fix typing of timesteps and init_timestep. 2024-06-25 11:31:52 -07:00
Ryan Dick
afaebdf151 Fix typing to reflect that the callback arg to latents_from_embeddings is never None. 2024-06-25 11:31:52 -07:00
Ryan Dick
d661517d94 Move seed above optional params. 2024-06-25 11:31:52 -07:00
Ryan Dick
82a69a54ac Simplify handling of AddsMaskGuidance, and fix some related type errors. 2024-06-25 11:31:52 -07:00
Ryan Dick
ffc28176fe Remove unused num_inference_steps. 2024-06-25 11:31:52 -07:00
Ryan Dick
d08e405017 Fix ControlNetModel type hint import source. 2024-06-25 11:31:52 -07:00
Lincoln Stein
b03073d888
[MM] Add support for probing and loading SDXL VAE checkpoint files (#6524) 
* add support for probing and loading SDXL VAE checkpoint files

* broaden regexp probe for SDXL VAEs

---------

Co-authored-by: Lincoln Stein <lstein@gmail.com>
 2024-06-20 02:57:27 +00:00
Ryan Dick
8e47e005a7 Tidy SilenceWarnings context manager: 
- Fix type errors
- Enable SilenceWarnings to be used as both a context manager and a decorator
- Remove duplicate implementation
- Check the initial verbosity on __enter__() rather than __init__()
 2024-06-18 15:06:22 -04:00
Brandon Rising
63a7e19dbf Run ruff 2024-06-18 10:38:29 -04:00
Brandon Rising
fbc5a8ec65 Ignore validation on improperly formatted hashes (pytest) 2024-06-18 10:38:29 -04:00
Brandon Rising
8ce6e4540e Run ruff 2024-06-18 10:38:29 -04:00
Brandon Rising
f14f377ede Update validator list 2024-06-18 10:38:29 -04:00
Brandon Rising
1925f83f5e Update validator list 2024-06-18 10:38:29 -04:00
Brandon Rising
3a5ad6d112 Update validator list 2024-06-18 10:38:29 -04:00
Brandon Rising
41a6bb45f3 Initial functionality 2024-06-18 10:38:29 -04:00
Lincoln Stein
a3cb5da130
Improve RAM<->VRAM memory copy performance in LoRA patching and elsewhere (#6490) 
* allow model patcher to optimize away the unpatching step when feasible

* remove lazy_offloading functionality

* allow model patcher to optimize away the unpatching step when feasible

* remove lazy_offloading functionality

* do not save original weights if there is a CPU copy of state dict

* Update invokeai/backend/model_manager/load/load_base.py

Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>

* documentation fixes requested during penultimate review

* add non-blocking=True parameters to several torch.nn.Module.to() calls, for slight performance increases

* fix ruff errors

* prevent crash on non-cuda-enabled systems

---------

Co-authored-by: Lincoln Stein <lstein@gmail.com>
Co-authored-by: Kent Keirsey <31807370+hipsterusername@users.noreply.github.com>
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
 2024-06-13 17:10:03 +00:00
psychedelicious
fde58ce0a3 Merge remote-tracking branch 'origin/main' into lstein/feat/simple-mm2-api 2024-06-07 14:23:41 +10:00
Lincoln Stein
f81b8bc9f6 add support for generic loading of diffusers directories 2024-06-07 13:54:30 +10:00
Lincoln Stein
2871676f79
LoRA patching optimization (#6439) 
* allow model patcher to optimize away the unpatching step when feasible

* remove lazy_offloading functionality

* allow model patcher to optimize away the unpatching step when feasible

* remove lazy_offloading functionality

* do not save original weights if there is a CPU copy of state dict

* Update invokeai/backend/model_manager/load/load_base.py

Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>

* documentation fixes added during penultimate review

---------

Co-authored-by: Lincoln Stein <lstein@gmail.com>
Co-authored-by: Kent Keirsey <31807370+hipsterusername@users.noreply.github.com>
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>
 2024-06-06 13:53:35 +00:00
psychedelicious
e7513f6088 docs(mm): add comment in move_model_to_device 2024-06-03 10:56:04 +10:00
Lincoln Stein
2276f327e5
Merge branch 'main' into lstein/feat/simple-mm2-api 2024-06-02 09:45:31 -04:00
Lincoln Stein
21a60af881
when unlocking models, offload_unlocked_models should prune to vram limit only (#6450) 
Co-authored-by: Lincoln Stein <lstein@gmail.com>
 2024-05-29 03:01:21 +00:00
Ryan Dick
829b9ad66b Add a callout about the hackiness of dropping tokens in the TextualInversionManager. 2024-05-28 05:11:54 -07:00
Ryan Dick
3aa1c8d3a8 Update TextualInversionManager for compatibility with the latest transformers release. See https://github.com/invoke-ai/InvokeAI/issues/6445. 2024-05-28 05:11:54 -07:00
Ryan Dick
994c61b67a Add docs to TextualInversionManager and improve types. No changes to functionality. 2024-05-28 05:11:54 -07:00
Lincoln Stein
34e1eb19f9 merge with main and resolve conflicts 2024-05-27 22:20:34 -04:00
Lincoln Stein
532f82cb97
Optimize RAM to VRAM transfer (#6312) 
* avoid copying model back from cuda to cpu

* handle models that don't have state dicts

* add assertions that models need a `device()` method

* do not rely on torch.nn.Module having the device() method

* apply all patches after model is on the execution device

* fix model patching in latents too

* log patched tokenizer

* closes #6375

---------

Co-authored-by: Lincoln Stein <lstein@gmail.com>
 2024-05-24 17:06:09 +00:00
Lincoln Stein
987ee704a1
Merge branch 'main' into lstein/feat/simple-mm2-api 2024-05-17 22:54:03 -04:00
Lincoln Stein
f29c406fed refactor model_install to work with refactored download queue 2024-05-13 22:49:15 -04:00
psychedelicious
93da75209c feat(nodes): use new blur_if_nsfw method 2024-05-14 07:23:38 +10:00
psychedelicious
9c819f0fd8 fix(nodes): fix nsfw checker model download 2024-05-14 07:23:38 +10:00
Lincoln Stein
0bf14c2830 add multifile_download() method to download service 2024-05-12 20:14:00 -06:00
blessedcoolant
da61396b1c cleanup: seamless unused older code cleanup 2024-05-13 08:11:08 +10:00
blessedcoolant
6c9fb617dc fix: fix seamless 2024-05-13 08:11:08 +10:00
Lincoln Stein
b48d4a049d bad implementation of diffusers folder download 2024-05-08 21:21:01 -07:00
Lincoln Stein
e9a20051bd refactor DWOpenPose and add type hints 2024-05-03 18:08:53 -04:00
Lincoln Stein
3b64e7a1fd
Merge branch 'main' into lstein/feat/simple-mm2-api 2024-05-02 21:20:35 -04:00
Lincoln Stein
49c84cd423
Merge branch 'main' into lstein/feat/simple-mm2-api 2024-04-30 18:13:42 -04:00
blessedcoolant
39ab4dd83e Merge branch 'main' into pr/6086 2024-05-01 00:37:06 +05:30
psychedelicious
2d7b8c2a1b fix(backend): do not round image dims to 64 in controlnet processor resize 
Rounding the dims results in control images that are subtly different than the input. We round to the nearest 8px later, there's no need to round now.
 2024-04-30 08:10:59 -04:00
psychedelicious
1fe90c357c feat(backend): lift managed model loading out of depthanything class 2024-04-29 08:56:00 +10:00
psychedelicious
fcb071f30c feat(backend): lift managed model loading out of lama class 2024-04-29 08:12:51 +10:00
Lincoln Stein
57c831442e fix safe_filename() on windows 2024-04-28 14:42:40 -04:00
Lincoln Stein
7c39929758 support VRAM caching of dict models that lack to() 2024-04-28 13:41:06 -04:00
Lincoln Stein
a26667d3ca make download and convert cache keys safe for filename length 2024-04-28 12:24:36 -04:00
Lincoln Stein
bb04f496e0 Merge branch 'main' into lstein/feat/simple-mm2-api 2024-04-28 11:33:26 -04:00
Lincoln Stein
70903ef057 refactor load_ckpt_from_url() 2024-04-28 11:33:23 -04:00
psychedelicious
241a1fdb57 feat(mm): support sdxl ckpt inpainting models 
There are only a couple SDXL inpainting models, and my tests indicate they are not as good as SD1.5 inpainting, but at least we support them now.

- Add the config file. This matches what is used in A1111. The only difference from the non-inpainting SDXL config is the number of in-channels.
- Update the legacy config maps to use this config file.
 2024-04-28 12:57:27 +10:00
Lincoln Stein
d72f272f16 Address change requests in first round of PR reviews. 
Pending:

- Move model install calls into model manager and create passthrus in invocation_context.
- Consider splitting load_model_from_url() into a call to get the path and a call to load the path.
 2024-04-24 23:53:30 -04:00
psychedelicious
6b0bf59682 feat(backend): update nms util to make blur/thresholding optional 2024-04-25 13:20:09 +10:00
blessedcoolant
260e24733f fix: update SDXL IP Adpater starter model to be ViT-H 2024-04-24 00:08:21 -04:00
blessedcoolant
6b394554e2 fix: update ip adapter starter models path 2024-04-24 08:48:25 +05:30
psychedelicious
a461537087 chore: ruff 2024-04-23 07:32:53 -04:00
psychedelicious
0aa5aadfe8 fix(mm): move variant to MainConfigBase 
shoulda been here all along
 2024-04-23 07:32:53 -04:00
Lincoln Stein
2b9f06dc4c
Re-enable app shutdown actions (#6244) 
* closes #6242

* only override sigINT during slow model scanning

* fix ruff formatting

---------

Co-authored-by: Lincoln Stein <lstein@gmail.com>
 2024-04-19 06:45:42 -04:00
Lincoln Stein
34cdfc61ab
Merge branch 'main' into lstein/feat/simple-mm2-api 2024-04-17 17:18:13 -04:00
blessedcoolant
6bab040d24 Merge branch 'main' into ip-adapter-style-comp 2024-04-16 21:14:06 +05:30
blessedcoolant
f46bbaf8c4 fix: make ip-adapter weights not be optional 2024-04-16 21:12:45 +05:30
blessedcoolant
d27907cc6d fix: entire reshaping block needs to be skipped 2024-04-16 04:29:53 +05:30
blessedcoolant
7ee3fef2db cleanup: better var names for the ip adapter weight collection block 2024-04-16 04:23:50 +05:30
blessedcoolant
a148c4322c fix: IP Adapter weights being incorrectly applied 
They were being overwritten rather than being appended
 2024-04-16 04:10:41 +05:30
blessedcoolant
5f6c6abf9c chore: change IPAdapterAttentionWeights to a dataclass 2024-04-15 23:38:55 +05:30
Lincoln Stein
470a39935c fix merge conflicts with main 2024-04-15 09:24:57 -04:00
Lincoln Stein
e93f4d632d
[util] Add generic torch device class (#6174) 
* introduce new abstraction layer for GPU devices

* add unit test for device abstraction

* fix ruff

* convert TorchDeviceSelect into a stateless class

* move logic to select context-specific execution device into context API

* add mock hardware environments to pytest

* remove dangling mocker fixture

* fix unit test for running on non-CUDA systems

* remove unimplemented get_execution_device() call

* remove autocast precision

* Multiple changes:

1. Remove TorchDeviceSelect.get_execution_device(), as well as calls to
   context.models.get_execution_device().
2. Rename TorchDeviceSelect to TorchDevice
3. Added back the legacy public API defined in `invocation_api`, including
   choose_precision().
4. Added a config file migration script to accommodate removal of precision=autocast.

* add deprecation warnings to choose_torch_device() and choose_precision()

* fix test crash

* remove app_config argument from choose_torch_device() and choose_torch_dtype()

---------

Co-authored-by: Lincoln Stein <lstein@gmail.com>
 2024-04-15 13:12:49 +00:00
Lincoln Stein
41b909cbe3 port dw_openpose, depth_anything, and lama processors to new model download scheme 2024-04-14 15:57:03 -04:00
blessedcoolant
8426f1e7b2 fix(experimental): Possible fix for conflict with regional embed length mismatch 
Pushing this so people can test it out and see if this needs to be handled in a different way.
 2024-04-14 12:19:19 +05:30
blessedcoolant
9cb0f63c44 refactor: fix a bunch of type issues in custom_attention 2024-04-13 14:17:25 +05:30
blessedcoolant
d4393e4170 chore: linter fixes 2024-04-13 12:14:45 +05:30
blessedcoolant
6ea183f0d4 wip: Initial Implementation IP Adapter Style & Comp Modes 2024-04-13 11:09:45 +05:30
Lincoln Stein
3a26c7bb9e fix merge conflicts 2024-04-12 00:58:11 -04:00
Lincoln Stein
df5ebdbc4f add invocation_context.load_ckpt_from_url() method 2024-04-12 00:55:21 -04:00
Lincoln Stein
651c0b39b1 clear cache on all exceptions 2024-04-12 07:19:16 +10:00
Lincoln Stein
46d23cd868 catch RunTimeError during model to() call rather than OutOfMemoryError 2024-04-12 07:19:16 +10:00
Lincoln Stein
579082ac10 [mm] clear the cache entry for a model that got an OOM during loading 2024-04-12 07:19:16 +10:00
psychedelicious
7bc77ddb40 fix(nodes): doubly-noised latents 
When using refiner with a mask (i.e. inpainting), we don't have noise provided as an input to the node.

This situation uniquely hits a code path that wasn't reviewed when gradient denoising was implemented.

That code path does two things wrong:
- It lerp'd the input latents. This was fixed in 5a1f4cb1ce.
- It added noise to the latents an extra time. This is fixed in this change.

We don't need to add noise in `latents_from_embeddings` because we do it just a lines later in `AddsMaskGuidance`.

- Remove the extraneous call to `add_noise`
- Make `seed` a required arg. We never call the function without seed anyways. If we refactor this in the future, it will be clearer that we need to look at how seed is handled.
- Move the call to create the noise to a deeper conditional, just before we call `AddsMaskGuidance`. The created noise tensor is now only used in that function, no need to create it every time.

Note: Whether or not having both noise and latents as inputs on the node is correct is a separate conversation. This change just fixes the issue with the current setup.
 2024-04-11 07:21:50 -04:00
Ryan Dick
f9af32a6d1 Fix the padding behavior when max-pooling regional IP-Adapter masks to mirror the downscaling behavior of SD and SDXL. Prior to this change, denoising with input latent dimensions that were not evenly divisible by 8 would raise an exception. 2024-04-09 16:50:43 -04:00
Ryan Dick
fba40eb1bd Fix the padding behavior when max-pooling regional prompt masks to mirror the downscaling behavior of SD and SDXL. Prior to this change, denoising with input latent dimensions that were not evenly divisible by 8 would raise an exception. 2024-04-09 16:50:43 -04:00
Ryan Dick
0c6dd32ece (minor) Fix IP-Adapter conditional logic in CustomAttnProcessor2_0. 2024-04-09 15:06:51 -04:00
Ryan Dick
0bdbfd4d1d Add support for IP-Adapter masks. 2024-04-09 15:06:51 -04:00
Ryan Dick
2e27ed5f3d Pass IP-Adapter scales through the cross_attn_kwargs pathway, since they are the same for all attention layers. This change also helps to prepare for adding IP-Adapter region masks. 2024-04-09 15:06:51 -04:00
Ryan Dick
4a828818da Remove support for Prompt-to-Prompt cross-attention control (aka .swap()). This feature is not widely used. It does not work with SDXL and is incompatible with IP-Adapter and regional prompting. The implementation is also intertwined with both text embedding and the UNet attention layers, resulting in a high maintenance burden. For all of these reasons, we have decided to drop support. 2024-04-09 10:57:02 -04:00
Ryan Dick
182810337c Add utility to_standard_float_mask(...) to convert various mask formats to a standardized format. 2024-04-09 08:12:12 -04:00
Ryan Dick
75ef473748 Pull the upstream changes from diffusers' AttnProcessor2_0 into CustomAttnProcessor2_0. This fixes a bug in CustomAttnProcessor2_0 that was being triggered when peft was not installed. The bug was present in a block of code that was previously copied from diffusers. The bug seems to have been introduced during diffusers' migration to PEFT for their LoRA handling. The upstream bug was fixed in 531e719163. 2024-04-09 08:12:12 -04:00
Ryan Dick
a78df8123f Update the diffusion logic to use the new regional prompting feature. 2024-04-09 08:12:12 -04:00
Ryan Dick
7ca677578e Create a UNetAttentionPatcher for patching UNet models with CustomAttnProcessor2_0 modules. 2024-04-09 08:12:12 -04:00
Ryan Dick
31c456c1e6 Update CustomAttention to support both IP-Adapters and regional prompting. 2024-04-09 08:12:12 -04:00
Ryan Dick
2ce79b61f5 Initialize a RegionalPromptAttnProcessor2_0 class by copying AttnProcessor2_0 from diffusers. 2024-04-09 08:12:12 -04:00
Ryan Dick
109e3f0e7f Add RegionalPromptData class for managing prompt region masks. 2024-04-09 08:12:12 -04:00
Ryan Dick
d1e45585d0 Add TextConditioningRegions to the TextConditioningData data structure. 2024-04-09 08:12:12 -04:00
Ryan Dick
aba023e0c5 Improve documentation of conditioning_data.py. 2024-04-09 08:12:12 -04:00
Ryan Dick
e354c29b52 Rename ConditioningData -> TextConditioningData. 2024-04-09 08:12:12 -04:00
Ryan Dick
a7f363e654 Split ip_adapter_conditioning out from ConditioningData. 2024-04-09 08:12:12 -04:00
Ryan Dick
9b2162e564 Remove scheduler_args from ConditioningData structure. 2024-04-09 08:12:12 -04:00
blessedcoolant
fd1f240853 fix: SDXL Refiner not working properly with Inpainting 2024-04-09 14:13:10 +10:00
psychedelicious
9ab6655491 feat(backend): clean up choose_precision 
- Allow user-defined precision on MPS.
- Use more explicit logic to handle all possible cases.
- Add comments.
- Remove the app_config args (they were effectively unused, just get the config using the singleton getter util)
 2024-04-07 09:41:05 -04:00
psychedelicious
4068e817d6 fix(mm): typing issues in model cache 2024-04-06 14:35:36 +11:00
psychedelicious
a09d705e4c fix(mm): remove vram check 
This check prematurely reports insufficient VRAM on Windows. See #6106 for details.
 2024-04-06 14:35:36 +11:00
Lincoln Stein
4571986c63 fix misplaced lock call 2024-04-05 14:32:18 +11:00
Lincoln Stein
812f10730f
adjust free vram calculation for models that will be removed by lazy offloading (#6150) 
Co-authored-by: Lincoln Stein <lstein@gmail.com>
 2024-04-04 22:51:12 -04:00