diff --git a/apps/shark_studio/api/sd.py b/apps/shark_studio/api/sd.py index 84bc9f9976..56db43a07c 100644 --- a/apps/shark_studio/api/sd.py +++ b/apps/shark_studio/api/sd.py @@ -26,7 +26,7 @@ get_resource_path, get_checkpoints_path, ) -from apps.shark_studio.modules.meta_model import SharkMetaModelBase +from apps.shark_studio.modules.meta_model import SharkMetaLoader from apps.shark_studio.modules.schedulers import get_schedulers from apps.shark_studio.modules.prompt_encoding import ( get_weighted_text_embeddings, @@ -80,23 +80,23 @@ "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))", ], }, - "vae_encode": { - "initializer": vae.export_vae_model, - "ireec_flags": [ - "--iree-flow-collapse-reduction-dims", - "--iree-opt-const-expr-hoisting=False", - "--iree-codegen-linalg-max-constant-fold-elements=9223372036854775807", - "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-detach-elementwise-from-named-ops,iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32},iree-linalg-ext-convert-conv2d-to-winograd))", - "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))", - ], - }, + # "vae_encode": { + # "initializer": vae.export_vae_model, + # "ireec_flags": [ + # "--iree-flow-collapse-reduction-dims", + # "--iree-opt-const-expr-hoisting=False", + # "--iree-codegen-linalg-max-constant-fold-elements=9223372036854775807", + # "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-detach-elementwise-from-named-ops,iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32},iree-linalg-ext-convert-conv2d-to-winograd))", + # "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-preprocessing-pad-linalg-ops{pad-size=16}))", + # ], + # }, "unet": { "initializer": unet.export_unet_model, "ireec_flags": [ "--iree-flow-collapse-reduction-dims", "--iree-opt-const-expr-hoisting=False", "--iree-codegen-linalg-max-constant-fold-elements=9223372036854775807", - "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32}))", + "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32},iree-linalg-ext-convert-conv2d-to-winograd))", ], }, "vae_decode": { @@ -138,7 +138,7 @@ def setup_shark( self.model_max_length = 77 self.batch_size = batch_size self.precision = precision - self.dtype = torch.float16 if precision == "fp16" else torch.float32 + # self.dtype = torch.float16 if precision == "fp16" else torch.float32 self.height = height self.width = width self.scheduler_obj = {} @@ -147,10 +147,10 @@ def setup_shark( "pipe": { "external_weights": "safetensors", }, - "clip": {"hf_model_name": self.base_model_id}, + "clip": { "hf_model_name": base_model_id }, "unet": { - "hf_model_name": self.base_model_id, - "unet_model": self.unet, + "hf_model_name": base_model_id, + "unet_model": unet.UnetModel(base_model_id, self.unet), "batch_size": batch_size, # "is_controlled": is_controlled, # "num_loras": num_loras, @@ -160,16 +160,16 @@ def setup_shark( "max_length": self.model_max_length, }, "vae_encode": { - "hf_model_name": self.base_model_id, - "vae_model": self.vae, + "hf_model_name": base_model_id, + "vae_model": vae.VaeModel(base_model_id, self.vae), "batch_size": batch_size, "height": height, "width": width, "precision": precision, }, "vae_decode": { - "hf_model_name": self.base_model_id, - "vae_model": self.vae, + "hf_model_name": base_model_id, + "vae_model": vae.VaeModel(base_model_id, self.vae), "batch_size": batch_size, "height": height, "width": width, @@ -179,7 +179,7 @@ def setup_shark( pipe_id_list = [ safe_name(self.base_model_id), str(batch_size), - str(self.compile_static_args["unet"]["max_length"]), + str(self.model_max_length), f"{str(height)}x{str(width)}", precision, ] @@ -200,7 +200,7 @@ def prepare_pipe(self, custom_weights, adapters, embeddings, is_img2img): self.schedulers = get_schedulers(self.base_model_id) self.weights_path = os.path.join( - get_checkpoints_path(), self.safe_name(self.base_model_id) + get_checkpoints_path(), self.shark_meta.safe_name(self.base_model_id) ) if not os.path.exists(self.weights_path): os.mkdir(self.weights_path) @@ -208,23 +208,23 @@ def prepare_pipe(self, custom_weights, adapters, embeddings, is_img2img): for model in adapters: self.model_map[model] = adapters[model] - for submodel in self.static_kwargs: + for submodel in self.compile_static_args: if custom_weights: custom_weights_params, _ = process_custom_pipe_weights(custom_weights) if submodel not in ["clip", "clip2"]: - self.static_kwargs[submodel][ + self.compile_static_args[submodel][ "external_weight_file" ] = custom_weights_params else: - self.static_kwargs[submodel]["external_weight_path"] = os.path.join( + self.compile_static_args[submodel]["external_weight_path"] = os.path.join( self.weights_path, submodel + ".safetensors" ) else: - self.static_kwargs[submodel]["external_weight_path"] = os.path.join( + self.compile_static_args[submodel]["external_weight_path"] = os.path.join( self.weights_path, submodel + ".safetensors" ) - self.get_compiled_map(pipe_id=self.pipe_id, static_kwargs=self.static_kwargs) + self.shark_meta.get_compiled_map(pipe_id=self.pipe_id, static_kwargs=self.compile_static_args) print("\n[LOG] Pipeline successfully prepared for runtime.") return @@ -288,17 +288,6 @@ def _get_pipeline_class( pipeline_cls = connected_pipeline_cls or pipeline_cls - # (SHARK): Monkey-patch in our classmethods to the auto-generated SD pipeline class. - pipeline_cls.setup_shark = setup_shark - pipeline_cls.prepare_pipe = prepare_pipe - - class shark_pipeline_class(SharkMetaModelBase, pipeline_cls): - def __getattribute__(self, __name: str) -> Any: - return super().__getattribute__(__name) - pass - - pipeline_cls = shark_pipeline_class - return pipeline_cls @@ -354,6 +343,7 @@ def device(self) -> torch.device: return torch.device("cpu") def register_modules(self, **kwargs): + print("sharkified") # import it here to avoid circular import diffusers_module = importlib.import_module("diffusers") pipelines = getattr(diffusers_module, "pipelines") @@ -777,8 +767,6 @@ def get_connected_passed_kwargs(prefix): ) # 7. Potentially add passed objects if expected - init_kwargs['device'] = device - init_kwargs['model_map'] = sd_model_map missing_modules = set(expected_modules) - set(init_kwargs.keys()) passed_modules = list(passed_class_obj.keys()) optional_modules = pipeline_class._optional_components @@ -790,12 +778,30 @@ def get_connected_passed_kwargs(prefix): raise ValueError( f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed." ) + + # 10. (SHARK): Monkey-patch in our classmethods to the auto-generated SD pipeline class. + pipeline_class.setup_shark = setup_shark + pipeline_class.prepare_pipe = prepare_pipe + pipeline_class.shark_meta = SharkMetaLoader(sd_model_map, device) + + # class shark_pipeline_class(pipeline_class): + # def __init__(self, pipeline_class, device, model_map, **init_kwargs): + # super().__init__(device, model_map, pipeline_class, **init_kwargs) + + + # def __getattribute__(cls, __name: str) -> Any: + # return super().__getattribute__(__name) + + #pipeline_class = shark_pipeline_class # 8. Instantiate the pipeline model = pipeline_class(**init_kwargs) # 9. Save where the model was instantiated from model.register_to_config(_name_or_path=pretrained_model_name_or_path) + + + return model @classmethod diff --git a/apps/shark_studio/modules/meta_model.py b/apps/shark_studio/modules/meta_model.py index 7da3a90a6a..cdb61bea0f 100644 --- a/apps/shark_studio/modules/meta_model.py +++ b/apps/shark_studio/modules/meta_model.py @@ -18,7 +18,7 @@ import os -class SharkMetaModelBase: +class SharkMetaLoader: # This class is a lightweight base for managing an # inference API class. It should provide methods for: # - compiling a set (model map) of torch IR modules @@ -30,37 +30,17 @@ def __init__( self, model_map: dict, device: str, - dtype: str = "f16", - import_mlir: bool = True, ): self.model_map = model_map self.pipe_map = {} self.triple = get_iree_target_triple(device) - self._device, self.device_id = clean_device_info(device) - self.import_mlir = import_mlir + self.device, self.device_id = clean_device_info(device) self.iree_module_dict = {} self.tmp_dir = get_resource_path(os.path.join("..", "shark_tmp")) if not os.path.exists(self.tmp_dir): os.mkdir(self.tmp_dir) self.tempfiles = {} self.pipe_vmfb_path = "" - self._dtype = dtype - - @property - def device(self): - return self._device - - @device.setter - def device(self, device_str): - self._device = device_str - - @property - def dtype(self): - return self._dtype - - @dtype.setter - def dtype(self, dtype_val): - self._dtype = dtype_val def get_compiled_map(self, pipe_id, static_kwargs, submodel="None", init_kwargs={}) -> None: # First checks whether we have .vmfbs precompiled, then populates the map @@ -96,32 +76,31 @@ def get_compiled_map(self, pipe_id, static_kwargs, submodel="None", init_kwargs= if key not in init_kwargs: init_kwargs[key] = static_kwargs["pipe"][key] self.import_torch_ir(submodel, init_kwargs) - self.get_compiled_map(pipe_id, submodel) + self.get_compiled_map(pipe_id, static_kwargs, submodel, init_kwargs) else: ireec_flags = ( self.model_map[submodel]["ireec_flags"] if "ireec_flags" in self.model_map[submodel] else [] ) - self.iree_module_dict[submodel] = get_iree_compiled_module( self.tempfiles[submodel], device=self.device, frontend="torch", mmap=True, - external_weight_file=self.get_io_params(submodel), + external_weight_file=self.get_io_params(submodel, static_kwargs), extra_args=ireec_flags, write_to=os.path.join(self.pipe_vmfb_path, submodel + ".vmfb"), ) return - def get_io_params(self, submodel): - if "external_weight_file" in self.static_kwargs[submodel]: + def get_io_params(self, submodel, static_kwargs: dict): + if "external_weight_file" in static_kwargs[submodel]: # we are using custom weights - weights_path = self.static_kwargs[submodel]["external_weight_file"] - elif "external_weight_path" in self.static_kwargs[submodel]: + weights_path = static_kwargs[submodel]["external_weight_file"] + elif "external_weight_path" in static_kwargs[submodel]: # we are using the default weights for the HF model - weights_path = self.static_kwargs[submodel]["external_weight_path"] + weights_path = static_kwargs[submodel]["external_weight_path"] else: # assume the torch IR contains the weights. weights_path = None @@ -143,6 +122,7 @@ def get_precompiled(self, pipe_id, submodel="None"): return def import_torch_ir(self, submodel, kwargs): + breakpoint() torch_ir = self.model_map[submodel]["initializer"]( **self.safe_dict(kwargs), compile_to="torch" ) diff --git a/apps/shark_studio/modules/pipeline.py b/apps/shark_studio/modules/pipeline.py deleted file mode 100644 index 5dee266b13..0000000000 --- a/apps/shark_studio/modules/pipeline.py +++ /dev/null @@ -1,206 +0,0 @@ -from msvcrt import kbhit -from shark.iree_utils.compile_utils import ( - get_iree_compiled_module, - load_vmfb_using_mmap, - clean_device_info, - get_iree_target_triple, -) -from apps.shark_studio.web.utils.file_utils import ( - get_checkpoints_path, - get_resource_path, -) -from apps.shark_studio.modules.shared_cmd_opts import ( - cmd_opts, -) -from iree import runtime as ireert -from pathlib import Path -import gc -import os - - -class SharkPipelineBase: - # This class is a lightweight base for managing an - # inference API class. It should provide methods for: - # - compiling a set (model map) of torch IR modules - # - preparing weights for an inference job - # - loading weights for an inference job - # - utilites like benchmarks, tests - - def __init__( - self, - model_map: dict, - base_model_id: str, - static_kwargs: dict, - device: str, - import_mlir: bool = True, - ): - self.model_map = model_map - self.pipe_map = {} - self.static_kwargs = static_kwargs - self.base_model_id = base_model_id - self.triple = get_iree_target_triple(device) - self.device, self.device_id = clean_device_info(device) - self.import_mlir = import_mlir - self.iree_module_dict = {} - self.tmp_dir = get_resource_path(os.path.join("..", "shark_tmp")) - if not os.path.exists(self.tmp_dir): - os.mkdir(self.tmp_dir) - self.tempfiles = {} - self.pipe_vmfb_path = "" - - def get_compiled_map(self, pipe_id, submodel="None", init_kwargs={}) -> None: - # First checks whether we have .vmfbs precompiled, then populates the map - # with the precompiled executables and fetches executables for the rest of the map. - # The weights aren't static here anymore so this function should be a part of pipeline - # initialization. As soon as you have a pipeline ID unique to your static torch IR parameters, - # and your model map is populated with any IR - unique model IDs and their static params, - # call this method to get the artifacts associated with your map. - self.pipe_id = self.safe_name(pipe_id) - self.pipe_vmfb_path = Path( - os.path.join(get_checkpoints_path(".."), self.pipe_id) - ) - self.pipe_vmfb_path.mkdir(parents=False, exist_ok=True) - if submodel == "None": - print("\n[LOG] Gathering any pre-compiled artifacts....") - for key in self.model_map: - self.get_compiled_map(pipe_id, submodel=key) - else: - self.pipe_map[submodel] = {} - self.get_precompiled(self.pipe_id, submodel) - ireec_flags = [] - if submodel in self.iree_module_dict: - return - elif "vmfb_path" in self.pipe_map[submodel]: - return - elif submodel not in self.tempfiles: - print( - f"\n[LOG] Tempfile for {submodel} not found. Fetching torch IR..." - ) - if submodel in self.static_kwargs: - init_kwargs = self.static_kwargs[submodel] - for key in self.static_kwargs["pipe"]: - if key not in init_kwargs: - init_kwargs[key] = self.static_kwargs["pipe"][key] - self.import_torch_ir(submodel, init_kwargs) - self.get_compiled_map(pipe_id, submodel) - else: - ireec_flags = ( - self.model_map[submodel]["ireec_flags"] - if "ireec_flags" in self.model_map[submodel] - else [] - ) - - weights_path = self.get_io_params(submodel) - - self.iree_module_dict[submodel] = get_iree_compiled_module( - self.tempfiles[submodel], - device=self.device, - frontend="torch", - mmap=True, - external_weight_file=weights_path, - extra_args=ireec_flags, - write_to=os.path.join(self.pipe_vmfb_path, submodel + ".vmfb"), - ) - return - - def get_io_params(self, submodel): - if "external_weight_file" in self.static_kwargs[submodel]: - # we are using custom weights - weights_path = self.static_kwargs[submodel]["external_weight_file"] - elif "external_weight_path" in self.static_kwargs[submodel]: - # we are using the default weights for the HF model - weights_path = self.static_kwargs[submodel]["external_weight_path"] - else: - # assume the torch IR contains the weights. - weights_path = None - return weights_path - - def get_precompiled(self, pipe_id, submodel="None"): - if submodel == "None": - for model in self.model_map: - self.get_precompiled(pipe_id, model) - vmfbs = [] - for dirpath, dirnames, filenames in os.walk(self.pipe_vmfb_path): - vmfbs.extend(filenames) - break - for file in vmfbs: - if submodel in file: - self.pipe_map[submodel]["vmfb_path"] = os.path.join( - self.pipe_vmfb_path, file - ) - return - - def import_torch_ir(self, submodel, kwargs): - torch_ir = self.model_map[submodel]["initializer"]( - **self.safe_dict(kwargs), compile_to="torch" - ) - if submodel == "clip": - # clip.export_clip_model returns (torch_ir, tokenizer) - torch_ir = torch_ir[0] - - self.tempfiles[submodel] = os.path.join( - self.tmp_dir, f"{submodel}.torch.tempfile" - ) - - with open(self.tempfiles[submodel], "w+") as f: - f.write(torch_ir) - del torch_ir - gc.collect() - return - - def load_submodels(self, submodels: list): - for submodel in submodels: - if submodel in self.iree_module_dict: - print(f"\n[LOG] {submodel} is ready for inference.") - continue - if "vmfb_path" in self.pipe_map[submodel]: - weights_path = self.get_io_params(submodel) - # print( - # f"\n[LOG] Loading .vmfb for {submodel} from {self.pipe_map[submodel]['vmfb_path']}" - # ) - self.iree_module_dict[submodel] = {} - ( - self.iree_module_dict[submodel]["vmfb"], - self.iree_module_dict[submodel]["config"], - self.iree_module_dict[submodel]["temp_file_to_unlink"], - ) = load_vmfb_using_mmap( - self.pipe_map[submodel]["vmfb_path"], - self.device, - device_idx=0, - rt_flags=[], - external_weight_file=weights_path, - ) - else: - self.get_compiled_map(self.pipe_id, submodel) - return - - def unload_submodels(self, submodels: list): - for submodel in submodels: - if submodel in self.iree_module_dict: - del self.iree_module_dict[submodel] - gc.collect() - return - - def run(self, submodel, inputs): - if not isinstance(inputs, list): - inputs = [inputs] - inp = [ - ireert.asdevicearray( - self.iree_module_dict[submodel]["config"].device, input - ) - for input in inputs - ] - return self.iree_module_dict[submodel]["vmfb"]["main"](*inp) - - def safe_name(self, name): - return name.replace("/", "_").replace("-", "_").replace("\\", "_") - - def safe_dict(self, kwargs: dict): - flat_args = {} - for i in kwargs: - if isinstance(kwargs[i], dict) and "pass_dict" not in kwargs[i]: - flat_args[i] = [kwargs[i][j] for j in kwargs[i]] - else: - flat_args[i] = kwargs[i] - - return flat_args diff --git a/apps/shark_studio/modules/shark_diffusers_pipe.py b/apps/shark_studio/modules/shark_diffusers_pipe.py new file mode 100644 index 0000000000..cd1d4f0951 --- /dev/null +++ b/apps/shark_studio/modules/shark_diffusers_pipe.py @@ -0,0 +1,2315 @@ +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. team. +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import fnmatch +import importlib +import inspect +import os +import re +import sys +import warnings +from dataclasses import dataclass +from pathlib import Path +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from huggingface_hub import ( + ModelCard, + create_repo, + hf_hub_download, + model_info, + snapshot_download, +) +from huggingface_hub.utils import validate_hf_hub_args +from packaging import version +from requests.exceptions import HTTPError +from tqdm.auto import tqdm + +import transformers +from transformers import PreTrainedModel +from transformers.utils import FLAX_WEIGHTS_NAME as TRANSFORMERS_FLAX_WEIGHTS_NAME +from transformers.utils import SAFE_WEIGHTS_NAME as TRANSFORMERS_SAFE_WEIGHTS_NAME +from transformers.utils import WEIGHTS_NAME as TRANSFORMERS_WEIGHTS_NAME + +# from ..utils import FLAX_WEIGHTS_NAME, ONNX_EXTERNAL_WEIGHTS_NAME, ONNX_WEIGHTS_NAME, PushToHubMixin + +INDEX_FILE = "diffusion_pytorch_model.bin" +CUSTOM_PIPELINE_FILE_NAME = "pipeline.py" +DUMMY_MODULES_FOLDER = "diffusers.utils" +TRANSFORMERS_DUMMY_MODULES_FOLDER = "transformers.utils" +CONNECTED_PIPES_KEYS = ["prior"] + + +logger = logging.get_logger(__name__) + + +LOADABLE_CLASSES = { + "diffusers": { + "ModelMixin": ["save_pretrained", "from_pretrained"], + "SchedulerMixin": ["save_pretrained", "from_pretrained"], + "DiffusionPipeline": ["save_pretrained", "from_pretrained"], + "OnnxRuntimeModel": ["save_pretrained", "from_pretrained"], + }, + "transformers": { + "PreTrainedTokenizer": ["save_pretrained", "from_pretrained"], + "PreTrainedTokenizerFast": ["save_pretrained", "from_pretrained"], + "PreTrainedModel": ["save_pretrained", "from_pretrained"], + "FeatureExtractionMixin": ["save_pretrained", "from_pretrained"], + "ProcessorMixin": ["save_pretrained", "from_pretrained"], + "ImageProcessingMixin": ["save_pretrained", "from_pretrained"], + }, + # "onnxruntime.training": { + # "ORTModule": ["save_pretrained", "from_pretrained"], + # }, +} + +ALL_IMPORTABLE_CLASSES = {} +for library in LOADABLE_CLASSES: + ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library]) + + +@dataclass +class ImagePipelineOutput(BaseOutput): + """ + Output class for image pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + + +@dataclass +class AudioPipelineOutput(BaseOutput): + """ + Output class for audio pipelines. + + Args: + audios (`np.ndarray`) + List of denoised audio samples of a NumPy array of shape `(batch_size, num_channels, sample_rate)`. + """ + + audios: np.ndarray + + +def is_safetensors_compatible(filenames, variant=None, passed_components=None) -> bool: + """ + Checking for safetensors compatibility: + - By default, all models are saved with the default pytorch serialization, so we use the list of default pytorch + files to know which safetensors files are needed. + - The model is safetensors compatible only if there is a matching safetensors file for every default pytorch file. + + Converting default pytorch serialized filenames to safetensors serialized filenames: + - For models from the diffusers library, just replace the ".bin" extension with ".safetensors" + - For models from the transformers library, the filename changes from "pytorch_model" to "model", and the ".bin" + extension is replaced with ".safetensors" + """ + pt_filenames = [] + + sf_filenames = set() + + passed_components = passed_components or [] + + for filename in filenames: + _, extension = os.path.splitext(filename) + + if ( + len(filename.split("/")) == 2 + and filename.split("/")[0] in passed_components + ): + continue + + if extension == ".bin": + pt_filenames.append(os.path.normpath(filename)) + elif extension == ".safetensors": + sf_filenames.add(os.path.normpath(filename)) + + for filename in pt_filenames: + # filename = 'foo/bar/baz.bam' -> path = 'foo/bar', filename = 'baz', extention = '.bam' + path, filename = os.path.split(filename) + filename, extension = os.path.splitext(filename) + + if filename.startswith("pytorch_model"): + filename = filename.replace("pytorch_model", "model") + else: + filename = filename + + expected_sf_filename = os.path.normpath(os.path.join(path, filename)) + expected_sf_filename = f"{expected_sf_filename}.safetensors" + if expected_sf_filename not in sf_filenames: + logger.warning(f"{expected_sf_filename} not found") + return False + + return True + + +def variant_compatible_siblings( + filenames, variant=None +) -> Union[List[os.PathLike], str]: + weight_names = [ + WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + FLAX_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + ONNX_EXTERNAL_WEIGHTS_NAME, + ] + + if is_transformers_available(): + weight_names += [ + TRANSFORMERS_WEIGHTS_NAME, + TRANSFORMERS_SAFE_WEIGHTS_NAME, + TRANSFORMERS_FLAX_WEIGHTS_NAME, + ] + + # model_pytorch, diffusion_model_pytorch, ... + weight_prefixes = [w.split(".")[0] for w in weight_names] + # .bin, .safetensors, ... + weight_suffixs = [w.split(".")[-1] for w in weight_names] + # -00001-of-00002 + transformers_index_format = r"\d{5}-of-\d{5}" + + if variant is not None: + # `diffusion_pytorch_model.fp16.bin` as well as `model.fp16-00001-of-00002.safetensors` + variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$" + ) + # `text_encoder/pytorch_model.bin.index.fp16.json` + variant_index_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.{variant}\.json$" + ) + + # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors` + non_variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$" + ) + # `text_encoder/pytorch_model.bin.index.json` + non_variant_index_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.json" + ) + + if variant is not None: + variant_weights = { + f for f in filenames if variant_file_re.match(f.split("/")[-1]) is not None + } + variant_indexes = { + f for f in filenames if variant_index_re.match(f.split("/")[-1]) is not None + } + variant_filenames = variant_weights | variant_indexes + else: + variant_filenames = set() + + non_variant_weights = { + f for f in filenames if non_variant_file_re.match(f.split("/")[-1]) is not None + } + non_variant_indexes = { + f for f in filenames if non_variant_index_re.match(f.split("/")[-1]) is not None + } + non_variant_filenames = non_variant_weights | non_variant_indexes + + # all variant filenames will be used by default + usable_filenames = set(variant_filenames) + + def convert_to_variant(filename): + if "index" in filename: + variant_filename = filename.replace("index", f"index.{variant}") + elif ( + re.compile(f"^(.*?){transformers_index_format}").match(filename) is not None + ): + variant_filename = f"{filename.split('-')[0]}.{variant}-{'-'.join(filename.split('-')[1:])}" + else: + variant_filename = ( + f"{filename.split('.')[0]}.{variant}.{filename.split('.')[1]}" + ) + return variant_filename + + for f in non_variant_filenames: + variant_filename = convert_to_variant(f) + if variant_filename not in usable_filenames: + usable_filenames.add(f) + + return usable_filenames, variant_filenames + + +@validate_hf_hub_args +def warn_deprecated_model_variant( + pretrained_model_name_or_path, token, variant, revision, model_filenames +): + info = model_info( + pretrained_model_name_or_path, + token=token, + revision=None, + ) + filenames = {sibling.rfilename for sibling in info.siblings} + comp_model_filenames, _ = variant_compatible_siblings(filenames, variant=revision) + comp_model_filenames = [ + ".".join(f.split(".")[:1] + f.split(".")[2:]) for f in comp_model_filenames + ] + + if set(model_filenames).issubset(set(comp_model_filenames)): + warnings.warn( + f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'` even though you can load it via `variant=`{revision}`. Loading model variants via `revision='{revision}'` is deprecated and will be removed in diffusers v1. Please use `variant='{revision}'` instead.", + FutureWarning, + ) + else: + warnings.warn( + f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'`. This behavior is deprecated and will be removed in diffusers v1. One should use `variant='{revision}'` instead. However, it appears that {pretrained_model_name_or_path} currently does not have the required variant filenames in the 'main' branch. \n The Diffusers team and community would be very grateful if you could open an issue: https://github.com/huggingface/diffusers/issues/new with the title '{pretrained_model_name_or_path} is missing {revision} files' so that the correct variant file can be added.", + FutureWarning, + ) + + +def _unwrap_model(model): + """Unwraps a model.""" + if is_compiled_module(model): + model = model._orig_mod + + if is_peft_available(): + from peft import PeftModel + + if isinstance(model, PeftModel): + model = model.base_model.model + + return model + + +def maybe_raise_or_warn( + library_name, + library, + class_name, + importable_classes, + passed_class_obj, + name, + is_pipeline_module, +): + """Simple helper method to raise or warn in case incorrect module has been passed""" + if not is_pipeline_module: + library = importlib.import_module(library_name) + class_obj = getattr(library, class_name) + class_candidates = { + c: getattr(library, c, None) for c in importable_classes.keys() + } + + expected_class_obj = None + for class_name, class_candidate in class_candidates.items(): + if class_candidate is not None and issubclass(class_obj, class_candidate): + expected_class_obj = class_candidate + + # Dynamo wraps the original model in a private class. + # I didn't find a public API to get the original class. + sub_model = passed_class_obj[name] + unwrapped_sub_model = _unwrap_model(sub_model) + model_cls = unwrapped_sub_model.__class__ + + if not issubclass(model_cls, expected_class_obj): + raise ValueError( + f"{passed_class_obj[name]} is of type: {model_cls}, but should be" + f" {expected_class_obj}" + ) + else: + logger.warning( + f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it" + " has the correct type" + ) + + +def get_class_obj_and_candidates( + library_name, + class_name, + importable_classes, + pipelines, + is_pipeline_module, + component_name=None, + cache_dir=None, +): + """Simple helper method to retrieve class object of module as well as potential parent class objects""" + component_folder = os.path.join(cache_dir, component_name) + + if is_pipeline_module: + pipeline_module = getattr(pipelines, library_name) + + class_obj = getattr(pipeline_module, class_name) + class_candidates = {c: class_obj for c in importable_classes.keys()} + elif os.path.isfile(os.path.join(component_folder, library_name + ".py")): + # load custom component + class_obj = get_class_from_dynamic_module( + component_folder, module_file=library_name + ".py", class_name=class_name + ) + class_candidates = {c: class_obj for c in importable_classes.keys()} + else: + # else we just import it from the library. + library = importlib.import_module(library_name) + + class_obj = getattr(library, class_name) + class_candidates = { + c: getattr(library, c, None) for c in importable_classes.keys() + } + + return class_obj, class_candidates + + +def _get_pipeline_class( + class_obj, + config, + load_connected_pipeline=False, + custom_pipeline=None, + repo_id=None, + hub_revision=None, + class_name=None, + cache_dir=None, + revision=None, +): + if custom_pipeline is not None: + if custom_pipeline.endswith(".py"): + path = Path(custom_pipeline) + # decompose into folder & file + file_name = path.name + custom_pipeline = path.parent.absolute() + elif repo_id is not None: + file_name = f"{custom_pipeline}.py" + custom_pipeline = repo_id + else: + file_name = CUSTOM_PIPELINE_FILE_NAME + + if repo_id is not None and hub_revision is not None: + # if we load the pipeline code from the Hub + # make sure to overwrite the `revison` + revision = hub_revision + + return get_class_from_dynamic_module( + custom_pipeline, + module_file=file_name, + class_name=class_name, + cache_dir=cache_dir, + revision=revision, + ) + + if class_obj != DiffusionPipeline: + return class_obj + + diffusers_module = importlib.import_module(class_obj.__module__.split(".")[0]) + class_name = config["_class_name"] + class_name = class_name[4:] if class_name.startswith("Flax") else class_name + + pipeline_cls = getattr(diffusers_module, class_name) + + if load_connected_pipeline: + from .auto_pipeline import _get_connected_pipeline + + connected_pipeline_cls = _get_connected_pipeline(pipeline_cls) + if connected_pipeline_cls is not None: + logger.info( + f"Loading connected pipeline {connected_pipeline_cls.__name__} instead of {pipeline_cls.__name__} as specified via `load_connected_pipeline=True`" + ) + else: + logger.info( + f"{pipeline_cls.__name__} has no connected pipeline class. Loading {pipeline_cls.__name__}." + ) + + pipeline_cls = connected_pipeline_cls or pipeline_cls + + return pipeline_cls + + +def load_sub_model( + library_name: str, + class_name: str, + importable_classes: List[Any], + pipelines: Any, + is_pipeline_module: bool, + pipeline_class: Any, + torch_dtype: torch.dtype, + provider: Any, + sess_options: Any, + device_map: Optional[Union[Dict[str, torch.device], str]], + max_memory: Optional[Dict[Union[int, str], Union[int, str]]], + offload_folder: Optional[Union[str, os.PathLike]], + offload_state_dict: bool, + model_variants: Dict[str, str], + name: str, + from_flax: bool, + variant: str, + low_cpu_mem_usage: bool, + cached_folder: Union[str, os.PathLike], + revision: str = None, +): + """Helper method to load the module `name` from `library_name` and `class_name`""" + # retrieve class candidates + class_obj, class_candidates = get_class_obj_and_candidates( + library_name, + class_name, + importable_classes, + pipelines, + is_pipeline_module, + component_name=name, + cache_dir=cached_folder, + ) + + load_method_name = None + # retrive load method name + for class_name, class_candidate in class_candidates.items(): + if class_candidate is not None and issubclass(class_obj, class_candidate): + load_method_name = importable_classes[class_name][1] + + # if load method name is None, then we have a dummy module -> raise Error + if load_method_name is None: + none_module = class_obj.__module__ + is_dummy_path = none_module.startswith( + DUMMY_MODULES_FOLDER + ) or none_module.startswith(TRANSFORMERS_DUMMY_MODULES_FOLDER) + if is_dummy_path and "dummy" in none_module: + # call class_obj for nice error message of missing requirements + class_obj() + + raise ValueError( + f"The component {class_obj} of {pipeline_class} cannot be loaded as it does not seem to have" + f" any of the loading methods defined in {ALL_IMPORTABLE_CLASSES}." + ) + + load_method = getattr(class_obj, load_method_name) + + # add kwargs to loading method + diffusers_module = importlib.import_module(__name__.split(".")[0]) + loading_kwargs = {} + if issubclass(class_obj, torch.nn.Module): + loading_kwargs["torch_dtype"] = torch_dtype + if issubclass(class_obj, diffusers_module.OnnxRuntimeModel): + loading_kwargs["provider"] = provider + loading_kwargs["sess_options"] = sess_options + + is_diffusers_model = issubclass(class_obj, diffusers_module.ModelMixin) + + if is_transformers_available(): + transformers_version = version.parse( + version.parse(transformers.__version__).base_version + ) + else: + transformers_version = "N/A" + + is_transformers_model = ( + is_transformers_available() + and issubclass(class_obj, PreTrainedModel) + and transformers_version >= version.parse("4.20.0") + ) + + # When loading a transformers model, if the device_map is None, the weights will be initialized as opposed to diffusers. + # To make default loading faster we set the `low_cpu_mem_usage=low_cpu_mem_usage` flag which is `True` by default. + # This makes sure that the weights won't be initialized which significantly speeds up loading. + if is_diffusers_model or is_transformers_model: + loading_kwargs["device_map"] = device_map + loading_kwargs["max_memory"] = max_memory + loading_kwargs["offload_folder"] = offload_folder + loading_kwargs["offload_state_dict"] = offload_state_dict + loading_kwargs["variant"] = model_variants.pop(name, None) + if from_flax: + loading_kwargs["from_flax"] = True + + # the following can be deleted once the minimum required `transformers` version + # is higher than 4.27 + if ( + is_transformers_model + and loading_kwargs["variant"] is not None + and transformers_version < version.parse("4.27.0") + ): + raise ImportError( + f"When passing `variant='{variant}'`, please make sure to upgrade your `transformers` version to at least 4.27.0.dev0" + ) + elif is_transformers_model and loading_kwargs["variant"] is None: + loading_kwargs.pop("variant") + + # if `from_flax` and model is transformer model, can currently not load with `low_cpu_mem_usage` + if not (from_flax and is_transformers_model): + loading_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage + else: + loading_kwargs["low_cpu_mem_usage"] = False + + # check if the module is in a subdirectory + if os.path.isdir(os.path.join(cached_folder, name)): + loaded_sub_model = load_method( + os.path.join(cached_folder, name), **loading_kwargs + ) + else: + # else load from the root directory + loaded_sub_model = load_method(cached_folder, **loading_kwargs) + + return loaded_sub_model + + +class DiffusionPipeline(ConfigMixin, PushToHubMixin): + r""" + Base class for all pipelines. + + [`DiffusionPipeline`] stores all components (models, schedulers, and processors) for diffusion pipelines and + provides methods for loading, downloading and saving models. It also includes methods to: + + - move all PyTorch modules to the device of your choice + - enable/disable the progress bar for the denoising iteration + + Class attributes: + + - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the + diffusion pipeline's components. + - **_optional_components** (`List[str]`) -- List of all optional components that don't have to be passed to the + pipeline to function (should be overridden by subclasses). + """ + + config_name = "model_index.json" + model_cpu_offload_seq = None + _optional_components = [] + _exclude_from_cpu_offload = [] + _load_connected_pipes = False + _is_onnx = False + + def register_modules(self, **kwargs): + # import it here to avoid circular import + diffusers_module = importlib.import_module(__name__.split(".")[0]) + pipelines = getattr(diffusers_module, "pipelines") + + for name, module in kwargs.items(): + # retrieve library + if ( + module is None + or isinstance(module, (tuple, list)) + and module[0] is None + ): + register_dict = {name: (None, None)} + else: + # register the config from the original module, not the dynamo compiled one + not_compiled_module = _unwrap_model(module) + + library = not_compiled_module.__module__.split(".")[0] + + # check if the module is a pipeline module + module_path_items = not_compiled_module.__module__.split(".") + pipeline_dir = ( + module_path_items[-2] if len(module_path_items) > 2 else None + ) + + path = not_compiled_module.__module__.split(".") + is_pipeline_module = pipeline_dir in path and hasattr( + pipelines, pipeline_dir + ) + + # if library is not in LOADABLE_CLASSES, then it is a custom module. + # Or if it's a pipeline module, then the module is inside the pipeline + # folder so we set the library to module name. + if is_pipeline_module: + library = pipeline_dir + elif library not in LOADABLE_CLASSES: + library = not_compiled_module.__module__ + + # retrieve class_name + class_name = not_compiled_module.__class__.__name__ + + register_dict = {name: (library, class_name)} + + # save model index config + self.register_to_config(**register_dict) + + # set models + setattr(self, name, module) + + def __setattr__(self, name: str, value: Any): + if name in self.__dict__ and hasattr(self.config, name): + # We need to overwrite the config if name exists in config + if isinstance(getattr(self.config, name), (tuple, list)): + if value is not None and self.config[name][0] is not None: + class_library_tuple = ( + value.__module__.split(".")[0], + value.__class__.__name__, + ) + else: + class_library_tuple = (None, None) + + self.register_to_config(**{name: class_library_tuple}) + else: + self.register_to_config(**{name: value}) + + super().__setattr__(name, value) + + def save_pretrained( + self, + save_directory: Union[str, os.PathLike], + safe_serialization: bool = True, + variant: Optional[str] = None, + push_to_hub: bool = False, + **kwargs, + ): + """ + Save all saveable variables of the pipeline to a directory. A pipeline variable can be saved and loaded if its + class implements both a save and loading method. The pipeline is easily reloaded using the + [`~DiffusionPipeline.from_pretrained`] class method. + + Arguments: + save_directory (`str` or `os.PathLike`): + Directory to save a pipeline to. Will be created if it doesn't exist. + safe_serialization (`bool`, *optional*, defaults to `True`): + Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`. + variant (`str`, *optional*): + If specified, weights are saved in the format `pytorch_model..bin`. + push_to_hub (`bool`, *optional*, defaults to `False`): + Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the + repository you want to push to with `repo_id` (will default to the name of `save_directory` in your + namespace). + kwargs (`Dict[str, Any]`, *optional*): + Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method. + """ + model_index_dict = dict(self.config) + model_index_dict.pop("_class_name", None) + model_index_dict.pop("_diffusers_version", None) + model_index_dict.pop("_module", None) + model_index_dict.pop("_name_or_path", None) + + if push_to_hub: + commit_message = kwargs.pop("commit_message", None) + private = kwargs.pop("private", False) + create_pr = kwargs.pop("create_pr", False) + token = kwargs.pop("token", None) + repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1]) + repo_id = create_repo( + repo_id, exist_ok=True, private=private, token=token + ).repo_id + + expected_modules, optional_kwargs = self._get_signature_keys(self) + + def is_saveable_module(name, value): + if name not in expected_modules: + return False + if name in self._optional_components and value[0] is None: + return False + return True + + model_index_dict = { + k: v for k, v in model_index_dict.items() if is_saveable_module(k, v) + } + for pipeline_component_name in model_index_dict.keys(): + sub_model = getattr(self, pipeline_component_name) + model_cls = sub_model.__class__ + + # Dynamo wraps the original model in a private class. + # I didn't find a public API to get the original class. + if is_compiled_module(sub_model): + sub_model = _unwrap_model(sub_model) + model_cls = sub_model.__class__ + + save_method_name = None + # search for the model's base class in LOADABLE_CLASSES + for library_name, library_classes in LOADABLE_CLASSES.items(): + if library_name in sys.modules: + library = importlib.import_module(library_name) + else: + logger.info( + f"{library_name} is not installed. Cannot save {pipeline_component_name} as {library_classes} from {library_name}" + ) + + for base_class, save_load_methods in library_classes.items(): + class_candidate = getattr(library, base_class, None) + if class_candidate is not None and issubclass( + model_cls, class_candidate + ): + # if we found a suitable base class in LOADABLE_CLASSES then grab its save method + save_method_name = save_load_methods[0] + break + if save_method_name is not None: + break + + if save_method_name is None: + logger.warn( + f"self.{pipeline_component_name}={sub_model} of type {type(sub_model)} cannot be saved." + ) + # make sure that unsaveable components are not tried to be loaded afterward + self.register_to_config(**{pipeline_component_name: (None, None)}) + continue + + save_method = getattr(sub_model, save_method_name) + + # Call the save method with the argument safe_serialization only if it's supported + save_method_signature = inspect.signature(save_method) + save_method_accept_safe = ( + "safe_serialization" in save_method_signature.parameters + ) + save_method_accept_variant = "variant" in save_method_signature.parameters + + save_kwargs = {} + if save_method_accept_safe: + save_kwargs["safe_serialization"] = safe_serialization + if save_method_accept_variant: + save_kwargs["variant"] = variant + + save_method( + os.path.join(save_directory, pipeline_component_name), **save_kwargs + ) + + # finally save the config + self.save_config(save_directory) + + if push_to_hub: + self._upload_folder( + save_directory, + repo_id, + token=token, + commit_message=commit_message, + create_pr=create_pr, + ) + + def to(self, *args, **kwargs): + r""" + Performs Pipeline dtype and/or device conversion. A torch.dtype and torch.device are inferred from the + arguments of `self.to(*args, **kwargs).` + + + + If the pipeline already has the correct torch.dtype and torch.device, then it is returned as is. Otherwise, + the returned pipeline is a copy of self with the desired torch.dtype and torch.device. + + + + + Here are the ways to call `to`: + + - `to(dtype, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified + [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype) + - `to(device, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified + [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) + - `to(device=None, dtype=None, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the + specified [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) and + [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype) + + Arguments: + dtype (`torch.dtype`, *optional*): + Returns a pipeline with the specified + [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype) + device (`torch.Device`, *optional*): + Returns a pipeline with the specified + [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) + silence_dtype_warnings (`str`, *optional*, defaults to `False`): + Whether to omit warnings if the target `dtype` is not compatible with the target `device`. + + Returns: + [`DiffusionPipeline`]: The pipeline converted to specified `dtype` and/or `dtype`. + """ + + torch_dtype = kwargs.pop("torch_dtype", None) + if torch_dtype is not None: + deprecate("torch_dtype", "0.27.0", "") + torch_device = kwargs.pop("torch_device", None) + if torch_device is not None: + deprecate("torch_device", "0.27.0", "") + + dtype_kwarg = kwargs.pop("dtype", None) + device_kwarg = kwargs.pop("device", None) + silence_dtype_warnings = kwargs.pop("silence_dtype_warnings", False) + + if torch_dtype is not None and dtype_kwarg is not None: + raise ValueError( + "You have passed both `torch_dtype` and `dtype` as a keyword argument. Please make sure to only pass `dtype`." + ) + + dtype = torch_dtype or dtype_kwarg + + if torch_device is not None and device_kwarg is not None: + raise ValueError( + "You have passed both `torch_device` and `device` as a keyword argument. Please make sure to only pass `device`." + ) + + device = torch_device or device_kwarg + + dtype_arg = None + device_arg = None + if len(args) == 1: + if isinstance(args[0], torch.dtype): + dtype_arg = args[0] + else: + device_arg = torch.device(args[0]) if args[0] is not None else None + elif len(args) == 2: + if isinstance(args[0], torch.dtype): + raise ValueError( + "When passing two arguments, make sure the first corresponds to `device` and the second to `dtype`." + ) + device_arg = torch.device(args[0]) if args[0] is not None else None + dtype_arg = args[1] + elif len(args) > 2: + raise ValueError( + "Please make sure to pass at most two arguments (`device` and `dtype`) `.to(...)`" + ) + + if dtype is not None and dtype_arg is not None: + raise ValueError( + "You have passed `dtype` both as an argument and as a keyword argument. Please only pass one of the two." + ) + + dtype = dtype or dtype_arg + + if device is not None and device_arg is not None: + raise ValueError( + "You have passed `device` both as an argument and as a keyword argument. Please only pass one of the two." + ) + + device = device or device_arg + + # throw warning if pipeline is in "offloaded"-mode but user tries to manually set to GPU. + def module_is_sequentially_offloaded(module): + if not is_accelerate_available() or is_accelerate_version("<", "0.14.0"): + return False + + return hasattr(module, "_hf_hook") and not isinstance( + module._hf_hook, + (accelerate.hooks.CpuOffload, accelerate.hooks.AlignDevicesHook), + ) + + def module_is_offloaded(module): + if not is_accelerate_available() or is_accelerate_version( + "<", "0.17.0.dev0" + ): + return False + + return hasattr(module, "_hf_hook") and isinstance( + module._hf_hook, accelerate.hooks.CpuOffload + ) + + # .to("cuda") would raise an error if the pipeline is sequentially offloaded, so we raise our own to make it clearer + pipeline_is_sequentially_offloaded = any( + module_is_sequentially_offloaded(module) + for _, module in self.components.items() + ) + if ( + pipeline_is_sequentially_offloaded + and device + and torch.device(device).type == "cuda" + ): + raise ValueError( + "It seems like you have activated sequential model offloading by calling `enable_sequential_cpu_offload`, but are now attempting to move the pipeline to GPU. This is not compatible with offloading. Please, move your pipeline `.to('cpu')` or consider removing the move altogether if you use sequential offloading." + ) + + # Display a warning in this case (the operation succeeds but the benefits are lost) + pipeline_is_offloaded = any( + module_is_offloaded(module) for _, module in self.components.items() + ) + if pipeline_is_offloaded and device and torch.device(device).type == "cuda": + logger.warning( + f"It seems like you have activated model offloading by calling `enable_model_cpu_offload`, but are now manually moving the pipeline to GPU. It is strongly recommended against doing so as memory gains from offloading are likely to be lost. Offloading automatically takes care of moving the individual components {', '.join(self.components.keys())} to GPU when needed. To make sure offloading works as expected, you should consider moving the pipeline back to CPU: `pipeline.to('cpu')` or removing the move altogether if you use offloading." + ) + + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + is_offloaded = pipeline_is_offloaded or pipeline_is_sequentially_offloaded + for module in modules: + is_loaded_in_8bit = ( + hasattr(module, "is_loaded_in_8bit") and module.is_loaded_in_8bit + ) + + if is_loaded_in_8bit and dtype is not None: + logger.warning( + f"The module '{module.__class__.__name__}' has been loaded in 8bit and conversion to {torch_dtype} is not yet supported. Module is still in 8bit precision." + ) + + if is_loaded_in_8bit and device is not None: + logger.warning( + f"The module '{module.__class__.__name__}' has been loaded in 8bit and moving it to {torch_dtype} via `.to()` is not yet supported. Module is still on {module.device}." + ) + else: + module.to(device, dtype) + + if ( + module.dtype == torch.float16 + and str(device) in ["cpu"] + and not silence_dtype_warnings + and not is_offloaded + ): + logger.warning( + "Pipelines loaded with `dtype=torch.float16` cannot run with `cpu` device. It" + " is not recommended to move them to `cpu` as running them will fail. Please make" + " sure to use an accelerator to run the pipeline in inference, due to the lack of" + " support for`float16` operations on this device in PyTorch. Please, remove the" + " `torch_dtype=torch.float16` argument, or use another device for inference." + ) + return self + + @property + def device(self) -> torch.device: + r""" + Returns: + `torch.device`: The torch device on which the pipeline is located. + """ + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + for module in modules: + return module.device + + return torch.device("cpu") + + @property + def dtype(self) -> torch.dtype: + r""" + Returns: + `torch.dtype`: The torch dtype on which the pipeline is located. + """ + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + for module in modules: + return module.dtype + + return torch.float32 + + @classmethod + @validate_hf_hub_args + def from_pretrained( + cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs + ): + r""" + Instantiate a PyTorch diffusion pipeline from pretrained pipeline weights. + + The pipeline is set in evaluation mode (`model.eval()`) by default. + + If you get the error message below, you need to finetune the weights for your downstream task: + + ``` + Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match: + - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated + You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. + ``` + + Parameters: + pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline + hosted on the Hub. + - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights + saved using + [`~DiffusionPipeline.save_pretrained`]. + torch_dtype (`str` or `torch.dtype`, *optional*): + Override the default `torch.dtype` and load the model with another dtype. If "auto" is passed, the + dtype is automatically derived from the model's weights. + custom_pipeline (`str`, *optional*): + + + + 🧪 This is an experimental feature and may change in the future. + + + + Can be either: + + - A string, the *repo id* (for example `hf-internal-testing/diffusers-dummy-pipeline`) of a custom + pipeline hosted on the Hub. The repository must contain a file called pipeline.py that defines + the custom pipeline. + - A string, the *file name* of a community pipeline hosted on GitHub under + [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file + names must match the file name and not the pipeline script (`clip_guided_stable_diffusion` + instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the + current main branch of GitHub. + - A path to a directory (`./my_pipeline_directory/`) containing a custom pipeline. The directory + must contain a file called `pipeline.py` that defines the custom pipeline. + + For more information on how to load and create custom pipelines, please have a look at [Loading and + Adding Custom + Pipelines](https://huggingface.co/docs/diffusers/using-diffusers/custom_pipeline_overview) + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory where a downloaded pretrained model configuration is cached if the standard cache + is not used. + resume_download (`bool`, *optional*, defaults to `False`): + Whether or not to resume downloading the model weights and configuration files. If set to `False`, any + incompletely downloaded files are deleted. + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + custom_revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id similar to + `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a + custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*): + A map that specifies where each submodule should go. It doesn’t need to be defined for each + parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the + same device. + + Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For + more information about each option see [designing a device + map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map). + max_memory (`Dict`, *optional*): + A dictionary device identifier for the maximum memory. Will default to the maximum memory available for + each GPU and the available CPU RAM if unset. + offload_folder (`str` or `os.PathLike`, *optional*): + The path to offload weights if device_map contains the value `"disk"`. + offload_state_dict (`bool`, *optional*): + If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if + the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True` + when there is some disk offload. + low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`): + Speed up model loading only loading the pretrained weights and not initializing the weights. This also + tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. + Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this + argument to `True` will raise an error. + use_safetensors (`bool`, *optional*, defaults to `None`): + If set to `None`, the safetensors weights are downloaded if they're available **and** if the + safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors + weights. If set to `False`, safetensors weights are not loaded. + use_onnx (`bool`, *optional*, defaults to `None`): + If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights + will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is + `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending + with `.onnx` and `.pb`. + kwargs (remaining dictionary of keyword arguments, *optional*): + Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline + class). The overwritten components are passed directly to the pipelines `__init__` method. See example + below for more information. + variant (`str`, *optional*): + Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when + loading `from_flax`. + + + + To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with + `huggingface-cli login`. + + + + Examples: + + ```py + >>> from diffusers import DiffusionPipeline + + >>> # Download pipeline from huggingface.co and cache. + >>> pipeline = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256") + + >>> # Download pipeline that requires an authorization token + >>> # For more information on access tokens, please refer to this section + >>> # of the documentation](https://huggingface.co/docs/hub/security-tokens) + >>> pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5") + + >>> # Use a different scheduler + >>> from diffusers import LMSDiscreteScheduler + + >>> scheduler = LMSDiscreteScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.scheduler = scheduler + ``` + """ + cache_dir = kwargs.pop("cache_dir", None) + resume_download = kwargs.pop("resume_download", False) + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + local_files_only = kwargs.pop("local_files_only", None) + token = kwargs.pop("token", None) + revision = kwargs.pop("revision", None) + from_flax = kwargs.pop("from_flax", False) + torch_dtype = kwargs.pop("torch_dtype", None) + custom_pipeline = kwargs.pop("custom_pipeline", None) + custom_revision = kwargs.pop("custom_revision", None) + provider = kwargs.pop("provider", None) + sess_options = kwargs.pop("sess_options", None) + device_map = kwargs.pop("device_map", None) + max_memory = kwargs.pop("max_memory", None) + offload_folder = kwargs.pop("offload_folder", None) + offload_state_dict = kwargs.pop("offload_state_dict", False) + low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT) + variant = kwargs.pop("variant", None) + use_safetensors = kwargs.pop("use_safetensors", None) + use_onnx = kwargs.pop("use_onnx", None) + load_connected_pipeline = kwargs.pop("load_connected_pipeline", False) + + # 1. Download the checkpoints and configs + # use snapshot download here to get it working from from_pretrained + if not os.path.isdir(pretrained_model_name_or_path): + if pretrained_model_name_or_path.count("/") > 1: + raise ValueError( + f'The provided pretrained_model_name_or_path "{pretrained_model_name_or_path}"' + " is neither a valid local path nor a valid repo id. Please check the parameter." + ) + cached_folder = cls.download( + pretrained_model_name_or_path, + cache_dir=cache_dir, + resume_download=resume_download, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + from_flax=from_flax, + use_safetensors=use_safetensors, + use_onnx=use_onnx, + custom_pipeline=custom_pipeline, + custom_revision=custom_revision, + variant=variant, + load_connected_pipeline=load_connected_pipeline, + **kwargs, + ) + else: + cached_folder = pretrained_model_name_or_path + + config_dict = cls.load_config(cached_folder) + + # pop out "_ignore_files" as it is only needed for download + config_dict.pop("_ignore_files", None) + + # 2. Define which model components should load variants + # We retrieve the information by matching whether variant + # model checkpoints exist in the subfolders + model_variants = {} + if variant is not None: + for folder in os.listdir(cached_folder): + folder_path = os.path.join(cached_folder, folder) + is_folder = os.path.isdir(folder_path) and folder in config_dict + variant_exists = is_folder and any( + p.split(".")[1].startswith(variant) for p in os.listdir(folder_path) + ) + if variant_exists: + model_variants[folder] = variant + + # 3. Load the pipeline class, if using custom module then load it from the hub + # if we load from explicit class, let's use it + custom_class_name = None + if os.path.isfile(os.path.join(cached_folder, f"{custom_pipeline}.py")): + custom_pipeline = os.path.join(cached_folder, f"{custom_pipeline}.py") + elif isinstance(config_dict["_class_name"], (list, tuple)) and os.path.isfile( + os.path.join(cached_folder, f"{config_dict['_class_name'][0]}.py") + ): + custom_pipeline = os.path.join( + cached_folder, f"{config_dict['_class_name'][0]}.py" + ) + custom_class_name = config_dict["_class_name"][1] + + pipeline_class = _get_pipeline_class( + cls, + config_dict, + load_connected_pipeline=load_connected_pipeline, + custom_pipeline=custom_pipeline, + class_name=custom_class_name, + cache_dir=cache_dir, + revision=custom_revision, + ) + + # DEPRECATED: To be removed in 1.0.0 + if ( + pipeline_class.__name__ == "StableDiffusionInpaintPipeline" + and version.parse( + version.parse(config_dict["_diffusers_version"]).base_version + ) + <= version.parse("0.5.1") + ): + from diffusers import ( + StableDiffusionInpaintPipeline, + StableDiffusionInpaintPipelineLegacy, + ) + + pipeline_class = StableDiffusionInpaintPipelineLegacy + + deprecation_message = ( + "You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the" + f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For" + " better inpainting results, we strongly suggest using Stable Diffusion's official inpainting" + " checkpoint: https://huggingface.co/runwayml/stable-diffusion-inpainting instead or adapting your" + f" checkpoint {pretrained_model_name_or_path} to the format of" + " https://huggingface.co/runwayml/stable-diffusion-inpainting. Note that we do not actively maintain" + " the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0." + ) + deprecate( + "StableDiffusionInpaintPipelineLegacy", + "1.0.0", + deprecation_message, + standard_warn=False, + ) + + # 4. Define expected modules given pipeline signature + # and define non-None initialized modules (=`init_kwargs`) + + # some modules can be passed directly to the init + # in this case they are already instantiated in `kwargs` + # extract them here + expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class) + passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs} + passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs} + + init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict( + config_dict, **kwargs + ) + + # define init kwargs and make sure that optional component modules are filtered out + init_kwargs = { + k: init_dict.pop(k) + for k in optional_kwargs + if k in init_dict and k not in pipeline_class._optional_components + } + init_kwargs = {**init_kwargs, **passed_pipe_kwargs} + + # remove `null` components + def load_module(name, value): + if value[0] is None: + return False + if name in passed_class_obj and passed_class_obj[name] is None: + return False + return True + + init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)} + + # Special case: safety_checker must be loaded separately when using `from_flax` + if ( + from_flax + and "safety_checker" in init_dict + and "safety_checker" not in passed_class_obj + ): + raise NotImplementedError( + "The safety checker cannot be automatically loaded when loading weights `from_flax`." + " Please, pass `safety_checker=None` to `from_pretrained`, and load the safety checker" + " separately if you need it." + ) + + # 5. Throw nice warnings / errors for fast accelerate loading + if len(unused_kwargs) > 0: + logger.warning( + f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored." + ) + + if low_cpu_mem_usage and not is_accelerate_available(): + low_cpu_mem_usage = False + logger.warning( + "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the" + " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install" + " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip" + " install accelerate\n```\n." + ) + + if device_map is not None and not is_torch_version(">=", "1.9.0"): + raise NotImplementedError( + "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set" + " `device_map=None`." + ) + + if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"): + raise NotImplementedError( + "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set" + " `low_cpu_mem_usage=False`." + ) + + if low_cpu_mem_usage is False and device_map is not None: + raise ValueError( + f"You cannot set `low_cpu_mem_usage` to False while using device_map={device_map} for loading and" + " dispatching. Please make sure to set `low_cpu_mem_usage=True`." + ) + + # import it here to avoid circular import + from diffusers import pipelines + + # 6. Load each module in the pipeline + for name, (library_name, class_name) in logging.tqdm( + init_dict.items(), desc="Loading pipeline components..." + ): + # 6.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names + class_name = class_name[4:] if class_name.startswith("Flax") else class_name + + # 6.2 Define all importable classes + is_pipeline_module = hasattr(pipelines, library_name) + importable_classes = ALL_IMPORTABLE_CLASSES + loaded_sub_model = None + + # 6.3 Use passed sub model or load class_name from library_name + if name in passed_class_obj: + # if the model is in a pipeline module, then we load it from the pipeline + # check that passed_class_obj has correct parent class + maybe_raise_or_warn( + library_name, + library, + class_name, + importable_classes, + passed_class_obj, + name, + is_pipeline_module, + ) + + loaded_sub_model = passed_class_obj[name] + else: + # load sub model + loaded_sub_model = load_sub_model( + library_name=library_name, + class_name=class_name, + importable_classes=importable_classes, + pipelines=pipelines, + is_pipeline_module=is_pipeline_module, + pipeline_class=pipeline_class, + torch_dtype=torch_dtype, + provider=provider, + sess_options=sess_options, + device_map=device_map, + max_memory=max_memory, + offload_folder=offload_folder, + offload_state_dict=offload_state_dict, + model_variants=model_variants, + name=name, + from_flax=from_flax, + variant=variant, + low_cpu_mem_usage=low_cpu_mem_usage, + cached_folder=cached_folder, + revision=revision, + ) + logger.info( + f"Loaded {name} as {class_name} from `{name}` subfolder of {pretrained_model_name_or_path}." + ) + + init_kwargs[name] = loaded_sub_model # UNet(...), # DiffusionSchedule(...) + + if pipeline_class._load_connected_pipes and os.path.isfile( + os.path.join(cached_folder, "README.md") + ): + modelcard = ModelCard.load(os.path.join(cached_folder, "README.md")) + connected_pipes = { + prefix: getattr(modelcard.data, prefix, [None])[0] + for prefix in CONNECTED_PIPES_KEYS + } + load_kwargs = { + "cache_dir": cache_dir, + "resume_download": resume_download, + "force_download": force_download, + "proxies": proxies, + "local_files_only": local_files_only, + "token": token, + "revision": revision, + "torch_dtype": torch_dtype, + "custom_pipeline": custom_pipeline, + "custom_revision": custom_revision, + "provider": provider, + "sess_options": sess_options, + "device_map": device_map, + "max_memory": max_memory, + "offload_folder": offload_folder, + "offload_state_dict": offload_state_dict, + "low_cpu_mem_usage": low_cpu_mem_usage, + "variant": variant, + "use_safetensors": use_safetensors, + } + + def get_connected_passed_kwargs(prefix): + connected_passed_class_obj = { + k.replace(f"{prefix}_", ""): w + for k, w in passed_class_obj.items() + if k.split("_")[0] == prefix + } + connected_passed_pipe_kwargs = { + k.replace(f"{prefix}_", ""): w + for k, w in passed_pipe_kwargs.items() + if k.split("_")[0] == prefix + } + + connected_passed_kwargs = { + **connected_passed_class_obj, + **connected_passed_pipe_kwargs, + } + return connected_passed_kwargs + + connected_pipes = { + prefix: DiffusionPipeline.from_pretrained( + repo_id, **load_kwargs.copy(), **get_connected_passed_kwargs(prefix) + ) + for prefix, repo_id in connected_pipes.items() + if repo_id is not None + } + + for prefix, connected_pipe in connected_pipes.items(): + # add connected pipes to `init_kwargs` with _, e.g. "prior_text_encoder" + init_kwargs.update( + { + "_".join([prefix, name]): component + for name, component in connected_pipe.components.items() + } + ) + + # 7. Potentially add passed objects if expected + missing_modules = set(expected_modules) - set(init_kwargs.keys()) + passed_modules = list(passed_class_obj.keys()) + optional_modules = pipeline_class._optional_components + if len(missing_modules) > 0 and missing_modules <= set( + passed_modules + optional_modules + ): + for module in missing_modules: + init_kwargs[module] = passed_class_obj.get(module, None) + elif len(missing_modules) > 0: + passed_modules = ( + set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) + - optional_kwargs + ) + raise ValueError( + f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed." + ) + + # 8. Instantiate the pipeline + model = pipeline_class(**init_kwargs) + + # 9. Save where the model was instantiated from + model.register_to_config(_name_or_path=pretrained_model_name_or_path) + return model + + @property + def name_or_path(self) -> str: + return getattr(self.config, "_name_or_path", None) + + @property + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + [`~DiffusionPipeline.enable_sequential_cpu_offload`] the execution device can only be inferred from + Accelerate's module hooks. + """ + for name, model in self.components.items(): + if ( + not isinstance(model, torch.nn.Module) + or name in self._exclude_from_cpu_offload + ): + continue + + if not hasattr(model, "_hf_hook"): + return self.device + for module in model.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def enable_model_cpu_offload( + self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda" + ): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward` + method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with + `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`. + + Arguments: + gpu_id (`int`, *optional*): + The ID of the accelerator that shall be used in inference. If not specified, it will default to 0. + device (`torch.Device` or `str`, *optional*, defaults to "cuda"): + The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will + default to "cuda". + """ + if self.model_cpu_offload_seq is None: + raise ValueError( + "Model CPU offload cannot be enabled because no `model_cpu_offload_seq` class attribute is set." + ) + + if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"): + from accelerate import cpu_offload_with_hook + else: + raise ImportError( + "`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher." + ) + + torch_device = torch.device(device) + device_index = torch_device.index + + if gpu_id is not None and device_index is not None: + raise ValueError( + f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}" + f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}" + ) + + # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0 + self._offload_gpu_id = ( + gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0) + ) + + device_type = torch_device.type + device = torch.device(f"{device_type}:{self._offload_gpu_id}") + + if self.device.type != "cpu": + self.to("cpu", silence_dtype_warnings=True) + device_mod = getattr(torch, self.device.type, None) + if hasattr(device_mod, "empty_cache") and device_mod.is_available(): + device_mod.empty_cache() # otherwise we don't see the memory savings (but they probably exist) + + all_model_components = { + k: v for k, v in self.components.items() if isinstance(v, torch.nn.Module) + } + + self._all_hooks = [] + hook = None + for model_str in self.model_cpu_offload_seq.split("->"): + model = all_model_components.pop(model_str, None) + if not isinstance(model, torch.nn.Module): + continue + + _, hook = cpu_offload_with_hook(model, device, prev_module_hook=hook) + self._all_hooks.append(hook) + + # CPU offload models that are not in the seq chain unless they are explicitly excluded + # these models will stay on CPU until maybe_free_model_hooks is called + # some models cannot be in the seq chain because they are iteratively called, such as controlnet + for name, model in all_model_components.items(): + if not isinstance(model, torch.nn.Module): + continue + + if name in self._exclude_from_cpu_offload: + model.to(device) + else: + _, hook = cpu_offload_with_hook(model, device) + self._all_hooks.append(hook) + + def maybe_free_model_hooks(self): + r""" + Function that offloads all components, removes all model hooks that were added when using + `enable_model_cpu_offload` and then applies them again. In case the model has not been offloaded this function + is a no-op. Make sure to add this function to the end of the `__call__` function of your pipeline so that it + functions correctly when applying enable_model_cpu_offload. + """ + if not hasattr(self, "_all_hooks") or len(self._all_hooks) == 0: + # `enable_model_cpu_offload` has not be called, so silently do nothing + return + + for hook in self._all_hooks: + # offload model and remove hook from model + hook.offload() + hook.remove() + + # make sure the model is in the same state as before calling it + self.enable_model_cpu_offload() + + def enable_sequential_cpu_offload( + self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda" + ): + r""" + Offloads all models to CPU using 🤗 Accelerate, significantly reducing memory usage. When called, the state + dicts of all `torch.nn.Module` components (except those in `self._exclude_from_cpu_offload`) are saved to CPU + and then moved to `torch.device('meta')` and loaded to GPU only when their specific submodule has its `forward` + method called. Offloading happens on a submodule basis. Memory savings are higher than with + `enable_model_cpu_offload`, but performance is lower. + + Arguments: + gpu_id (`int`, *optional*): + The ID of the accelerator that shall be used in inference. If not specified, it will default to 0. + device (`torch.Device` or `str`, *optional*, defaults to "cuda"): + The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will + default to "cuda". + """ + if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"): + from accelerate import cpu_offload + else: + raise ImportError( + "`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher" + ) + + torch_device = torch.device(device) + device_index = torch_device.index + + if gpu_id is not None and device_index is not None: + raise ValueError( + f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}" + f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}" + ) + + # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0 + self._offload_gpu_id = ( + gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0) + ) + + device_type = torch_device.type + device = torch.device(f"{device_type}:{self._offload_gpu_id}") + + if self.device.type != "cpu": + self.to("cpu", silence_dtype_warnings=True) + device_mod = getattr(torch, self.device.type, None) + if hasattr(device_mod, "empty_cache") and device_mod.is_available(): + device_mod.empty_cache() # otherwise we don't see the memory savings (but they probably exist) + + for name, model in self.components.items(): + if not isinstance(model, torch.nn.Module): + continue + + if name in self._exclude_from_cpu_offload: + model.to(device) + else: + # make sure to offload buffers if not all high level weights + # are of type nn.Module + offload_buffers = len(model._parameters) > 0 + cpu_offload(model, device, offload_buffers=offload_buffers) + + @classmethod + @validate_hf_hub_args + def download(cls, pretrained_model_name, **kwargs) -> Union[str, os.PathLike]: + r""" + Download and cache a PyTorch diffusion pipeline from pretrained pipeline weights. + + Parameters: + pretrained_model_name (`str` or `os.PathLike`, *optional*): + A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline + hosted on the Hub. + custom_pipeline (`str`, *optional*): + Can be either: + + - A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained + pipeline hosted on the Hub. The repository must contain a file called `pipeline.py` that defines + the custom pipeline. + + - A string, the *file name* of a community pipeline hosted on GitHub under + [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file + names must match the file name and not the pipeline script (`clip_guided_stable_diffusion` + instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the + current `main` branch of GitHub. + + - A path to a *directory* (`./my_pipeline_directory/`) containing a custom pipeline. The directory + must contain a file called `pipeline.py` that defines the custom pipeline. + + + + 🧪 This is an experimental feature and may change in the future. + + + + For more information on how to load and create custom pipelines, take a look at [How to contribute a + community pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/contribute_pipeline). + + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + resume_download (`bool`, *optional*, defaults to `False`): + Whether or not to resume downloading the model weights and configuration files. If set to `False`, any + incompletely downloaded files are deleted. + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + custom_revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id similar to + `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a + custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you're downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + variant (`str`, *optional*): + Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when + loading `from_flax`. + use_safetensors (`bool`, *optional*, defaults to `None`): + If set to `None`, the safetensors weights are downloaded if they're available **and** if the + safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors + weights. If set to `False`, safetensors weights are not loaded. + use_onnx (`bool`, *optional*, defaults to `False`): + If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights + will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is + `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending + with `.onnx` and `.pb`. + trust_remote_code (`bool`, *optional*, defaults to `False`): + Whether or not to allow for custom pipelines and components defined on the Hub in their own files. This + option should only be set to `True` for repositories you trust and in which you have read the code, as + it will execute code present on the Hub on your local machine. + + Returns: + `os.PathLike`: + A path to the downloaded pipeline. + + + + To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with + `huggingface-cli login`. + + + + """ + cache_dir = kwargs.pop("cache_dir", None) + resume_download = kwargs.pop("resume_download", False) + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + local_files_only = kwargs.pop("local_files_only", None) + token = kwargs.pop("token", None) + revision = kwargs.pop("revision", None) + from_flax = kwargs.pop("from_flax", False) + custom_pipeline = kwargs.pop("custom_pipeline", None) + custom_revision = kwargs.pop("custom_revision", None) + variant = kwargs.pop("variant", None) + use_safetensors = kwargs.pop("use_safetensors", None) + use_onnx = kwargs.pop("use_onnx", None) + load_connected_pipeline = kwargs.pop("load_connected_pipeline", False) + trust_remote_code = kwargs.pop("trust_remote_code", False) + + allow_pickle = False + if use_safetensors is None: + use_safetensors = True + allow_pickle = True + + allow_patterns = None + ignore_patterns = None + + model_info_call_error: Optional[Exception] = None + if not local_files_only: + try: + info = model_info(pretrained_model_name, token=token, revision=revision) + except HTTPError as e: + logger.warn( + f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache." + ) + local_files_only = True + model_info_call_error = ( + e # save error to reraise it if model is not cached locally + ) + + if not local_files_only: + config_file = hf_hub_download( + pretrained_model_name, + cls.config_name, + cache_dir=cache_dir, + revision=revision, + proxies=proxies, + force_download=force_download, + resume_download=resume_download, + token=token, + ) + + config_dict = cls._dict_from_json_file(config_file) + ignore_filenames = config_dict.pop("_ignore_files", []) + + # retrieve all folder_names that contain relevant files + folder_names = [ + k + for k, v in config_dict.items() + if isinstance(v, list) and k != "_class_name" + ] + + filenames = {sibling.rfilename for sibling in info.siblings} + model_filenames, variant_filenames = variant_compatible_siblings( + filenames, variant=variant + ) + + diffusers_module = importlib.import_module(__name__.split(".")[0]) + pipelines = getattr(diffusers_module, "pipelines") + + # optionally create a custom component <> custom file mapping + custom_components = {} + for component in folder_names: + module_candidate = config_dict[component][0] + + if module_candidate is None or not isinstance(module_candidate, str): + continue + + # We compute candidate file path on the Hub. Do not use `os.path.join`. + candidate_file = f"{component}/{module_candidate}.py" + + if candidate_file in filenames: + custom_components[component] = module_candidate + elif module_candidate not in LOADABLE_CLASSES and not hasattr( + pipelines, module_candidate + ): + raise ValueError( + f"{candidate_file} as defined in `model_index.json` does not exist in {pretrained_model_name} and is not a module in 'diffusers/pipelines'." + ) + + if len(variant_filenames) == 0 and variant is not None: + deprecation_message = ( + f"You are trying to load the model files of the `variant={variant}`, but no such modeling files are available." + f"The default model files: {model_filenames} will be loaded instead. Make sure to not load from `variant={variant}`" + "if such variant modeling files are not available. Doing so will lead to an error in v0.24.0 as defaulting to non-variant" + "modeling files is deprecated." + ) + deprecate( + "no variant default", + "0.24.0", + deprecation_message, + standard_warn=False, + ) + + # remove ignored filenames + model_filenames = set(model_filenames) - set(ignore_filenames) + variant_filenames = set(variant_filenames) - set(ignore_filenames) + + # if the whole pipeline is cached we don't have to ping the Hub + if revision in DEPRECATED_REVISION_ARGS and version.parse( + version.parse(__version__).base_version + ) >= version.parse("0.22.0"): + warn_deprecated_model_variant( + pretrained_model_name, token, variant, revision, model_filenames + ) + + model_folder_names = { + os.path.split(f)[0] + for f in model_filenames + if os.path.split(f)[0] in folder_names + } + + custom_class_name = None + if custom_pipeline is None and isinstance( + config_dict["_class_name"], (list, tuple) + ): + custom_pipeline = config_dict["_class_name"][0] + custom_class_name = config_dict["_class_name"][1] + + # all filenames compatible with variant will be added + allow_patterns = list(model_filenames) + + # allow all patterns from non-model folders + # this enables downloading schedulers, tokenizers, ... + allow_patterns += [ + f"{k}/*" for k in folder_names if k not in model_folder_names + ] + # add custom component files + allow_patterns += [f"{k}/{f}.py" for k, f in custom_components.items()] + # add custom pipeline file + allow_patterns += ( + [f"{custom_pipeline}.py"] + if f"{custom_pipeline}.py" in filenames + else [] + ) + # also allow downloading config.json files with the model + allow_patterns += [ + os.path.join(k, "config.json") for k in model_folder_names + ] + + allow_patterns += [ + SCHEDULER_CONFIG_NAME, + CONFIG_NAME, + cls.config_name, + CUSTOM_PIPELINE_FILE_NAME, + ] + + load_pipe_from_hub = ( + custom_pipeline is not None and f"{custom_pipeline}.py" in filenames + ) + load_components_from_hub = len(custom_components) > 0 + + if load_pipe_from_hub and not trust_remote_code: + raise ValueError( + f"The repository for {pretrained_model_name} contains custom code in {custom_pipeline}.py which must be executed to correctly " + f"load the model. You can inspect the repository content at https://hf.co/{pretrained_model_name}/blob/main/{custom_pipeline}.py.\n" + f"Please pass the argument `trust_remote_code=True` to allow custom code to be run." + ) + + if load_components_from_hub and not trust_remote_code: + raise ValueError( + f"The repository for {pretrained_model_name} contains custom code in {'.py, '.join([os.path.join(k, v) for k,v in custom_components.items()])} which must be executed to correctly " + f"load the model. You can inspect the repository content at {', '.join([f'https://hf.co/{pretrained_model_name}/{k}/{v}.py' for k,v in custom_components.items()])}.\n" + f"Please pass the argument `trust_remote_code=True` to allow custom code to be run." + ) + + # retrieve passed components that should not be downloaded + pipeline_class = _get_pipeline_class( + cls, + config_dict, + load_connected_pipeline=load_connected_pipeline, + custom_pipeline=custom_pipeline, + repo_id=pretrained_model_name if load_pipe_from_hub else None, + hub_revision=revision, + class_name=custom_class_name, + cache_dir=cache_dir, + revision=custom_revision, + ) + expected_components, _ = cls._get_signature_keys(pipeline_class) + passed_components = [k for k in expected_components if k in kwargs] + + if ( + use_safetensors + and not allow_pickle + and not is_safetensors_compatible( + model_filenames, + variant=variant, + passed_components=passed_components, + ) + ): + raise EnvironmentError( + f"Could not find the necessary `safetensors` weights in {model_filenames} (variant={variant})" + ) + if from_flax: + ignore_patterns = ["*.bin", "*.safetensors", "*.onnx", "*.pb"] + elif use_safetensors and is_safetensors_compatible( + model_filenames, variant=variant, passed_components=passed_components + ): + ignore_patterns = ["*.bin", "*.msgpack"] + + use_onnx = use_onnx if use_onnx is not None else pipeline_class._is_onnx + if not use_onnx: + ignore_patterns += ["*.onnx", "*.pb"] + + safetensors_variant_filenames = { + f for f in variant_filenames if f.endswith(".safetensors") + } + safetensors_model_filenames = { + f for f in model_filenames if f.endswith(".safetensors") + } + if ( + len(safetensors_variant_filenames) > 0 + and safetensors_model_filenames != safetensors_variant_filenames + ): + logger.warn( + f"\nA mixture of {variant} and non-{variant} filenames will be loaded.\nLoaded {variant} filenames:\n[{', '.join(safetensors_variant_filenames)}]\nLoaded non-{variant} filenames:\n[{', '.join(safetensors_model_filenames - safetensors_variant_filenames)}\nIf this behavior is not expected, please check your folder structure." + ) + else: + ignore_patterns = ["*.safetensors", "*.msgpack"] + + use_onnx = use_onnx if use_onnx is not None else pipeline_class._is_onnx + if not use_onnx: + ignore_patterns += ["*.onnx", "*.pb"] + + bin_variant_filenames = { + f for f in variant_filenames if f.endswith(".bin") + } + bin_model_filenames = {f for f in model_filenames if f.endswith(".bin")} + if ( + len(bin_variant_filenames) > 0 + and bin_model_filenames != bin_variant_filenames + ): + logger.warn( + f"\nA mixture of {variant} and non-{variant} filenames will be loaded.\nLoaded {variant} filenames:\n[{', '.join(bin_variant_filenames)}]\nLoaded non-{variant} filenames:\n[{', '.join(bin_model_filenames - bin_variant_filenames)}\nIf this behavior is not expected, please check your folder structure." + ) + + # Don't download any objects that are passed + allow_patterns = [ + p + for p in allow_patterns + if not (len(p.split("/")) == 2 and p.split("/")[0] in passed_components) + ] + + if pipeline_class._load_connected_pipes: + allow_patterns.append("README.md") + + # Don't download index files of forbidden patterns either + ignore_patterns = ignore_patterns + [ + f"{i}.index.*json" for i in ignore_patterns + ] + + re_ignore_pattern = [ + re.compile(fnmatch.translate(p)) for p in ignore_patterns + ] + re_allow_pattern = [ + re.compile(fnmatch.translate(p)) for p in allow_patterns + ] + + expected_files = [ + f for f in filenames if not any(p.match(f) for p in re_ignore_pattern) + ] + expected_files = [ + f for f in expected_files if any(p.match(f) for p in re_allow_pattern) + ] + + snapshot_folder = Path(config_file).parent + pipeline_is_cached = all( + (snapshot_folder / f).is_file() for f in expected_files + ) + + if pipeline_is_cached and not force_download: + # if the pipeline is cached, we can directly return it + # else call snapshot_download + return snapshot_folder + + user_agent = {"pipeline_class": cls.__name__} + if custom_pipeline is not None and not custom_pipeline.endswith(".py"): + user_agent["custom_pipeline"] = custom_pipeline + + # download all allow_patterns - ignore_patterns + try: + cached_folder = snapshot_download( + pretrained_model_name, + cache_dir=cache_dir, + resume_download=resume_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + allow_patterns=allow_patterns, + ignore_patterns=ignore_patterns, + user_agent=user_agent, + ) + + # retrieve pipeline class from local file + cls_name = cls.load_config( + os.path.join(cached_folder, "model_index.json") + ).get("_class_name", None) + cls_name = ( + cls_name[4:] + if isinstance(cls_name, str) and cls_name.startswith("Flax") + else cls_name + ) + + diffusers_module = importlib.import_module(__name__.split(".")[0]) + pipeline_class = ( + getattr(diffusers_module, cls_name, None) + if isinstance(cls_name, str) + else None + ) + + if pipeline_class is not None and pipeline_class._load_connected_pipes: + modelcard = ModelCard.load(os.path.join(cached_folder, "README.md")) + connected_pipes = sum( + [getattr(modelcard.data, k, []) for k in CONNECTED_PIPES_KEYS], [] + ) + for connected_pipe_repo_id in connected_pipes: + download_kwargs = { + "cache_dir": cache_dir, + "resume_download": resume_download, + "force_download": force_download, + "proxies": proxies, + "local_files_only": local_files_only, + "token": token, + "variant": variant, + "use_safetensors": use_safetensors, + } + DiffusionPipeline.download( + connected_pipe_repo_id, **download_kwargs + ) + + return cached_folder + + except FileNotFoundError: + # Means we tried to load pipeline with `local_files_only=True` but the files have not been found in local cache. + # This can happen in two cases: + # 1. If the user passed `local_files_only=True` => we raise the error directly + # 2. If we forced `local_files_only=True` when `model_info` failed => we raise the initial error + if model_info_call_error is None: + # 1. user passed `local_files_only=True` + raise + else: + # 2. we forced `local_files_only=True` when `model_info` failed + raise EnvironmentError( + f"Cannot load model {pretrained_model_name}: model is not cached locally and an error occured" + " while trying to fetch metadata from the Hub. Please check out the root cause in the stacktrace" + " above." + ) from model_info_call_error + + @classmethod + def _get_signature_keys(cls, obj): + parameters = inspect.signature(obj.__init__).parameters + required_parameters = { + k: v for k, v in parameters.items() if v.default == inspect._empty + } + optional_parameters = set( + {k for k, v in parameters.items() if v.default != inspect._empty} + ) + expected_modules = set(required_parameters.keys()) - {"self"} + + optional_names = list(optional_parameters) + for name in optional_names: + if name in cls._optional_components: + expected_modules.add(name) + optional_parameters.remove(name) + + return expected_modules, optional_parameters + + @property + def components(self) -> Dict[str, Any]: + r""" + The `self.components` property can be useful to run different pipelines with the same weights and + configurations without reallocating additional memory. + + Returns (`dict`): + A dictionary containing all the modules needed to initialize the pipeline. + + Examples: + + ```py + >>> from diffusers import ( + ... StableDiffusionPipeline, + ... StableDiffusionImg2ImgPipeline, + ... StableDiffusionInpaintPipeline, + ... ) + + >>> text2img = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5") + >>> img2img = StableDiffusionImg2ImgPipeline(**text2img.components) + >>> inpaint = StableDiffusionInpaintPipeline(**text2img.components) + ``` + """ + expected_modules, optional_parameters = self._get_signature_keys(self) + components = { + k: getattr(self, k) + for k in self.config.keys() + if not k.startswith("_") and k not in optional_parameters + } + + if set(components.keys()) != expected_modules: + raise ValueError( + f"{self} has been incorrectly initialized or {self.__class__} is incorrectly implemented. Expected" + f" {expected_modules} to be defined, but {components.keys()} are defined." + ) + + return components + + @staticmethod + def numpy_to_pil(images): + """ + Convert a NumPy image or a batch of images to a PIL image. + """ + return numpy_to_pil(images) + + def progress_bar(self, iterable=None, total=None): + if not hasattr(self, "_progress_bar_config"): + self._progress_bar_config = {} + elif not isinstance(self._progress_bar_config, dict): + raise ValueError( + f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}." + ) + + if iterable is not None: + return tqdm(iterable, **self._progress_bar_config) + elif total is not None: + return tqdm(total=total, **self._progress_bar_config) + else: + raise ValueError("Either `total` or `iterable` has to be defined.") + + def set_progress_bar_config(self, **kwargs): + self._progress_bar_config = kwargs + + def enable_xformers_memory_efficient_attention( + self, attention_op: Optional[Callable] = None + ): + r""" + Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/). When this + option is enabled, you should observe lower GPU memory usage and a potential speed up during inference. Speed + up during training is not guaranteed. + + + + ⚠️ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes + precedent. + + + + Parameters: + attention_op (`Callable`, *optional*): + Override the default `None` operator for use as `op` argument to the + [`memory_efficient_attention()`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention) + function of xFormers. + + Examples: + + ```py + >>> import torch + >>> from diffusers import DiffusionPipeline + >>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp + + >>> pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16) + >>> pipe = pipe.to("cuda") + >>> pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp) + >>> # Workaround for not accepting attention shape using VAE for Flash Attention + >>> pipe.vae.enable_xformers_memory_efficient_attention(attention_op=None) + ``` + """ + self.set_use_memory_efficient_attention_xformers(True, attention_op) + + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/). + """ + self.set_use_memory_efficient_attention_xformers(False) + + def set_use_memory_efficient_attention_xformers( + self, valid: bool, attention_op: Optional[Callable] = None + ) -> None: + # Recursively walk through all the children. + # Any children which exposes the set_use_memory_efficient_attention_xformers method + # gets the message + def fn_recursive_set_mem_eff(module: torch.nn.Module): + if hasattr(module, "set_use_memory_efficient_attention_xformers"): + module.set_use_memory_efficient_attention_xformers(valid, attention_op) + + for child in module.children(): + fn_recursive_set_mem_eff(child) + + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + for module in modules: + fn_recursive_set_mem_eff(module) + + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. When this option is enabled, the attention module splits the input tensor + in slices to compute attention in several steps. For more than one attention head, the computation is performed + sequentially over each head. This is useful to save some memory in exchange for a small speed decrease. + + + + ⚠️ Don't enable attention slicing if you're already using `scaled_dot_product_attention` (SDPA) from PyTorch + 2.0 or xFormers. These attention computations are already very memory efficient so you won't need to enable + this function. If you enable attention slicing with SDPA or xFormers, it can lead to serious slow downs! + + + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + `"max"`, maximum amount of memory will be saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + + Examples: + + ```py + >>> import torch + >>> from diffusers import StableDiffusionPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained( + ... "runwayml/stable-diffusion-v1-5", + ... torch_dtype=torch.float16, + ... use_safetensors=True, + ... ) + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> pipe.enable_attention_slicing() + >>> image = pipe(prompt).images[0] + ``` + """ + self.set_attention_slice(slice_size) + + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously called, attention is + computed in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def set_attention_slice(self, slice_size: Optional[int]): + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [ + m + for m in modules + if isinstance(m, torch.nn.Module) and hasattr(m, "set_attention_slice") + ] + + for module in modules: + module.set_attention_slice(slice_size) diff --git a/apps/shark_studio/tests/diffusers_pipeline_test.py b/apps/shark_studio/tests/diffusers_pipeline_test.py index 421d8a5457..d4c71e0daf 100644 --- a/apps/shark_studio/tests/diffusers_pipeline_test.py +++ b/apps/shark_studio/tests/diffusers_pipeline_test.py @@ -16,19 +16,18 @@ class SDBaseAPITest(unittest.TestCase): def testPipeSimple(self): pipe = SharkDiffusionPipeline.from_pretrained( - pretrained_model_name_or_path="hf-internal-testing/tiny-stable-diffusion-torch", + pretrained_model_name_or_path="stabilityai/stable-diffusion-2-1", device="vulkan", - torch_dtype=torch.float32, - ) + torch_dtype=torch.float16, + ).to(torch.float16) pipe.setup_shark( - base_model_id="hf-internal-testing/tiny-stable-diffusion-torch", + base_model_id="stabilityai/stable-diffusion-2-1", height=512, width=512, batch_size=1, - precision="f32", + precision="f16", device="vulkan", ) - pipe.prepare_pipe( custom_weights="", adapters=[],