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

Merge branch 'refactor/nodes-on-generator' of github.com:invoke-ai/InvokeAI into refactor/nodes-on-generator

Browse Source
This commit is contained in:
Lincoln Stein 2023-03-10 19:36:40 -05:00
commit fe75b95464
9 changed files with 30 additions and 136 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
docs/installation/020_INSTALL_MANUAL.md
View File

@ -148,7 +148,7 @@ manager, please follow these steps:
=== "CUDA (NVidia)"
```bash
pip install InvokeAI[xformers] --use-pep517 --extra-index-url https://download.pytorch.org/whl/cu117
pip install "InvokeAI[xformers]" --use-pep517 --extra-index-url https://download.pytorch.org/whl/cu117
```
=== "ROCm (AMD)"

12
invokeai/backend/generate.py
View File

@ -495,18 +495,6 @@ class Generate:
torch.cuda.reset_peak_memory_stats()
results = list()
init_image = None
mask_image = None
try:
if (
self.free_gpu_mem
and self.model.cond_stage_model.device != self.model.device
):
self.model.cond_stage_model.device = self.model.device
self.model.cond_stage_model.to(self.model.device)
except AttributeError:
pass
try:
uc, c, extra_conditioning_info = get_uc_and_c_and_ec(

2
invokeai/backend/model_management/convert_ckpt_to_diffusers.py
View File

@ -1274,7 +1274,7 @@ def load_pipeline_from_original_stable_diffusion_ckpt(
tokenizer=tokenizer,
unet=unet.to(precision),
scheduler=scheduler,
safety_checker=safety_checker.to(precision),
safety_checker=None if return_generator_pipeline else safety_checker.to(precision),
feature_extractor=feature_extractor,
)
else:

41
invokeai/backend/model_management/model_manager.py
View File

@ -108,7 +108,7 @@ class ModelManager(object):
if model_name in self.models:
requested_model = self.models[model_name]["model"]
print(f">> Retrieving model {model_name} from system RAM cache")
self.models[model_name]["model"] = self._model_from_cpu(requested_model)
requested_model.ready()
width = self.models[model_name]["width"]
height = self.models[model_name]["height"]
hash = self.models[model_name]["hash"]
@ -503,7 +503,7 @@ class ModelManager(object):
print(f">> Offloading {model_name} to CPU")
model = self.models[model_name]["model"]
self.models[model_name]["model"] = self._model_to_cpu(model)
model.offload_all()
gc.collect()
if self._has_cuda():
@ -1048,43 +1048,6 @@ class ModelManager(object):
self.stack.remove(model_name)
self.models.pop(model_name, None)
def _model_to_cpu(self, model):
if self.device == CPU_DEVICE:
return model
if isinstance(model, StableDiffusionGeneratorPipeline):
model.offload_all()
return model
model.cond_stage_model.device = CPU_DEVICE
model.to(CPU_DEVICE)
for submodel in ("first_stage_model", "cond_stage_model", "model"):
try:
getattr(model, submodel).to(CPU_DEVICE)
except AttributeError:
pass
return model
def _model_from_cpu(self, model):
if self.device == CPU_DEVICE:
return model
if isinstance(model, StableDiffusionGeneratorPipeline):
model.ready()
return model
model.to(self.device)
model.cond_stage_model.device = self.device
for submodel in ("first_stage_model", "cond_stage_model", "model"):
try:
getattr(model, submodel).to(self.device)
except AttributeError:
pass
return model
def _pop_oldest_model(self):
"""
Remove the first element of the FIFO, which ought

1
invokeai/backend/prompting/__init__.py
View File

@ -3,7 +3,6 @@ Initialization file for invokeai.backend.prompting
"""
from .conditioning import (
get_prompt_structure,
get_tokenizer,
get_tokens_for_prompt_object,
get_uc_and_c_and_ec,
split_weighted_subprompts,

58
invokeai/backend/prompting/conditioning.py
View File

@ -7,7 +7,7 @@ get_uc_and_c_and_ec() get the conditioned and unconditioned latent, an
"""
import re
from typing import Any, Optional, Union
from typing import Optional, Union
from compel import Compel
from compel.prompt_parser import (
@ -17,7 +17,6 @@ from compel.prompt_parser import (
Fragment,
PromptParser,
)
from transformers import CLIPTextModel, CLIPTokenizer
from invokeai.backend.globals import Globals
@ -25,36 +24,6 @@ from ..stable_diffusion import InvokeAIDiffuserComponent
from ..util import torch_dtype
def get_tokenizer(model) -> CLIPTokenizer:
# TODO remove legacy ckpt fallback handling
return (
getattr(model, "tokenizer", None) # diffusers
or model.cond_stage_model.tokenizer
) # ldm
def get_text_encoder(model) -> Any:
# TODO remove legacy ckpt fallback handling
return getattr(
model, "text_encoder", None
) or UnsqueezingLDMTransformer( # diffusers
model.cond_stage_model.transformer
) # ldm
class UnsqueezingLDMTransformer:
def __init__(self, ldm_transformer):
self.ldm_transformer = ldm_transformer
@property
def device(self):
return self.ldm_transformer.device
def __call__(self, *args, **kwargs):
insufficiently_unsqueezed_tensor = self.ldm_transformer(*args, **kwargs)
return insufficiently_unsqueezed_tensor.unsqueeze(0)
def get_uc_and_c_and_ec(
prompt_string, model, log_tokens=False, skip_normalize_legacy_blend=False
):
@ -64,13 +33,13 @@ def get_uc_and_c_and_ec(
prompt_string
)
tokenizer = get_tokenizer(model)
text_encoder = get_text_encoder(model)
tokenizer = model.tokenizer
compel = Compel(
tokenizer=tokenizer,
text_encoder=text_encoder,
text_encoder=model.text_encoder,
textual_inversion_manager=model.textual_inversion_manager,
dtype_for_device_getter=torch_dtype,
truncate_long_prompts=False
)
# get rid of any newline characters
@ -82,12 +51,12 @@ def get_uc_and_c_and_ec(
legacy_blend = try_parse_legacy_blend(
positive_prompt_string, skip_normalize_legacy_blend
)
positive_prompt: FlattenedPrompt | Blend
positive_prompt: Union[FlattenedPrompt, Blend]
if legacy_blend is not None:
positive_prompt = legacy_blend
else:
positive_prompt = Compel.parse_prompt_string(positive_prompt_string)
negative_prompt: FlattenedPrompt | Blend = Compel.parse_prompt_string(
negative_prompt: Union[FlattenedPrompt, Blend] = Compel.parse_prompt_string(
negative_prompt_string
)
@ -96,6 +65,7 @@ def get_uc_and_c_and_ec(
c, options = compel.build_conditioning_tensor_for_prompt_object(positive_prompt)
uc, _ = compel.build_conditioning_tensor_for_prompt_object(negative_prompt)
[c, uc] = compel.pad_conditioning_tensors_to_same_length([c, uc])
tokens_count = get_max_token_count(tokenizer, positive_prompt)
@ -116,12 +86,12 @@ def get_prompt_structure(
legacy_blend = try_parse_legacy_blend(
positive_prompt_string, skip_normalize_legacy_blend
)
positive_prompt: FlattenedPrompt | Blend
positive_prompt: Union[FlattenedPrompt, Blend]
if legacy_blend is not None:
positive_prompt = legacy_blend
else:
positive_prompt = Compel.parse_prompt_string(positive_prompt_string)
negative_prompt: FlattenedPrompt | Blend = Compel.parse_prompt_string(
negative_prompt: Union[FlattenedPrompt, Blend] = Compel.parse_prompt_string(
negative_prompt_string
)
@ -129,7 +99,7 @@ def get_prompt_structure(
def get_max_token_count(
tokenizer, prompt: Union[FlattenedPrompt, Blend], truncate_if_too_long=True
tokenizer, prompt: Union[FlattenedPrompt, Blend], truncate_if_too_long=False
) -> int:
if type(prompt) is Blend:
blend: Blend = prompt
@ -245,7 +215,7 @@ def log_tokenization_for_prompt_object(
)
def log_tokenization_for_text(text, tokenizer, display_label=None):
def log_tokenization_for_text(text, tokenizer, display_label=None, truncate_if_too_long=False):
"""shows how the prompt is tokenized
# usually tokens have '</w>' to indicate end-of-word,
# but for readability it has been replaced with ' '
@ -260,11 +230,11 @@ def log_tokenization_for_text(text, tokenizer, display_label=None):
token = tokens[i].replace("</w>", " ")
# alternate color
s = (usedTokens % 6) + 1
if i < tokenizer.model_max_length:
if truncate_if_too_long and i >= tokenizer.model_max_length:
discarded = discarded + f"\x1b[0;3{s};40m{token}"
else:
tokenized = tokenized + f"\x1b[0;3{s};40m{token}"
usedTokens += 1
else: # over max token length
discarded = discarded + f"\x1b[0;3{s};40m{token}"
if usedTokens > 0:
print(f'\n>> [TOKENLOG] Tokens {display_label or ""} ({usedTokens}):')

45
invokeai/backend/stable_diffusion/diffusers_pipeline.py
View File

@ -54,16 +54,6 @@ class PipelineIntermediateState:
attention_map_saver: Optional[AttentionMapSaver] = None
# copied from configs/stable-diffusion/v1-inference.yaml
_default_personalization_config_params = dict(
placeholder_strings=["*"],
initializer_wods=["sculpture"],
per_image_tokens=False,
num_vectors_per_token=1,
progressive_words=False,
)
@dataclass
class AddsMaskLatents:
"""Add the channels required for inpainting model input.
@ -175,7 +165,7 @@ def image_resized_to_grid_as_tensor(
:param normalize: scale the range to [-1, 1] instead of [0, 1]
:param multiple_of: resize the input so both dimensions are a multiple of this
"""
w, h = trim_to_multiple_of(*image.size)
w, h = trim_to_multiple_of(*image.size, multiple_of=multiple_of)
transformation = T.Compose(
[
T.Resize((h, w), T.InterpolationMode.LANCZOS),
@ -290,10 +280,10 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
scheduler ([`SchedulerMixin`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
safety_checker ([`StableDiffusionSafetyChecker`]):
Classification module that estimates whether generated images could be considered offsensive or harmful.
Classification module that estimates whether generated images could be considered offensive or harmful.
Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
feature_extractor ([`CLIPFeatureExtractor`]):
Model that extracts features from generated images to be used as inputs for the `safety_checker`.
@ -436,11 +426,11 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
"""
Ready this pipeline's models.
i.e. pre-load them to the GPU if appropriate.
i.e. preload them to the GPU if appropriate.
"""
self._model_group.ready()
def to(self, torch_device: Optional[Union[str, torch.device]] = None):
def to(self, torch_device: Optional[Union[str, torch.device]] = None, silence_dtype_warnings=False):
# overridden method; types match the superclass.
if torch_device is None:
return self
@ -917,20 +907,6 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
device=self._model_group.device_for(self.unet),
)
@property
def cond_stage_model(self):
return self.embeddings_provider
@torch.inference_mode()
def _tokenize(self, prompt: Union[str, List[str]]):
return self.tokenizer(
prompt,
padding="max_length",
max_length=self.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
@property
def channels(self) -> int:
"""Compatible with DiffusionWrapper"""
@ -942,11 +918,10 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
return super().decode_latents(latents)
def debug_latents(self, latents, msg):
from invokeai.backend.image_util import debug_image
with torch.inference_mode():
from ldm.util import debug_image
decoded = self.numpy_to_pil(self.decode_latents(latents))
for i, img in enumerate(decoded):
debug_image(
img, f"latents {msg} {i+1}/{len(decoded)}", debug_status=True
)
for i, img in enumerate(decoded):
debug_image(
img, f"latents {msg} {i+1}/{len(decoded)}", debug_status=True
)

3
invokeai/backend/web/invoke_ai_web_server.py
View File

@ -29,7 +29,6 @@ from ..image_util import PngWriter, retrieve_metadata
from ...frontend.merge.merge_diffusers import merge_diffusion_models
from ..prompting import (
get_prompt_structure,
get_tokenizer,
get_tokens_for_prompt_object,
)
from ..stable_diffusion import PipelineIntermediateState
@ -1274,7 +1273,7 @@ class InvokeAIWebServer:
None
if type(parsed_prompt) is Blend
else get_tokens_for_prompt_object(
get_tokenizer(self.generate.model), parsed_prompt
self.generate.model.tokenizer, parsed_prompt
)
)
attention_maps_image_base64_url = (

2
pyproject.toml
View File

@ -38,7 +38,7 @@ dependencies = [
"albumentations",
"click",
"clip_anytorch", # replacing "clip @ https://github.com/openai/CLIP/archive/eaa22acb90a5876642d0507623e859909230a52d.zip",
"compel==0.1.7",
"compel==0.1.10",
"datasets",
"diffusers[torch]~=0.14",
"dnspython==2.2.1",