# coding=utf-8

# Copyright 2025 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 copy

import functools

import inspect

import itertools

import json

import os

import re

import shutil

import tempfile

from collections import OrderedDict

from contextlib import ExitStack, contextmanager

from functools import wraps

from pathlib import Path

from typing import Any, Callable, ContextManager, Type

import safetensors

import torch

import torch.utils.checkpoint

from huggingface_hub import DDUFEntry, create_repo, split_torch_state_dict_into_shards

from huggingface_hub.utils import validate_hf_hub_args

from torch import Tensor, nn

from typing_extensions import Self

from .. import __version__

from ..quantizers import DiffusersAutoQuantizer, DiffusersQuantizer

from ..quantizers.quantization_config import QuantizationMethod

from ..utils import (

CONFIG_NAME,

FLAX_WEIGHTS_NAME,

HF_ENABLE_PARALLEL_LOADING,

SAFE_WEIGHTS_INDEX_NAME,

SAFETENSORS_WEIGHTS_NAME,

WEIGHTS_INDEX_NAME,

WEIGHTS_NAME,

_add_variant,

_get_checkpoint_shard_files,

_get_model_file,

deprecate,

is_accelerate_available,

is_bitsandbytes_available,

is_bitsandbytes_version,

is_peft_available,

is_torch_version,

logging,

)

from ..utils.distributed_utils import is_torch_dist_rank_zero

from ..utils.hub_utils import PushToHubMixin, load_or_create_model_card, populate_model_card

from ..utils.torch_utils import empty_device_cache

from ._modeling_parallel import ContextParallelConfig, ContextParallelModelPlan, ParallelConfig

from .model_loading_utils import (

_caching_allocator_warmup,

_determine_device_map,

_expand_device_map,

_fetch_index_file,

_fetch_index_file_legacy,

_load_shard_file,

_load_shard_files_with_threadpool,

load_state_dict,

)

class ContextManagers:

"""

Wrapper for `contextlib.ExitStack` which enters a collection of context managers. Adaptation of `ContextManagers`

in the `fastcore` library.

"""

def __init__(self, context_managers: list[ContextManager]):

self.context_managers = context_managers

self.stack = ExitStack()

def __enter__(self):

for context_manager in self.context_managers:

self.stack.enter_context(context_manager)

def __exit__(self, *args, **kwargs):

self.stack.__exit__(*args, **kwargs)

logger = logging.get_logger(__name__)

_REGEX_SHARD = re.compile(r"(.*?)-\d{5}-of-\d{5}")

TORCH_INIT_FUNCTIONS = {

"uniform_": nn.init.uniform_,

"normal_": nn.init.normal_,

"trunc_normal_": nn.init.trunc_normal_,

"constant_": nn.init.constant_,

"xavier_uniform_": nn.init.xavier_uniform_,

"xavier_normal_": nn.init.xavier_normal_,

"kaiming_uniform_": nn.init.kaiming_uniform_,

"kaiming_normal_": nn.init.kaiming_normal_,

"uniform": nn.init.uniform,

"normal": nn.init.normal,

"xavier_uniform": nn.init.xavier_uniform,

"xavier_normal": nn.init.xavier_normal,

"kaiming_uniform": nn.init.kaiming_uniform,

"kaiming_normal": nn.init.kaiming_normal,

}

if is_torch_version(">=", "1.9.0"):

_LOW_CPU_MEM_USAGE_DEFAULT = True

else:

_LOW_CPU_MEM_USAGE_DEFAULT = False

if is_accelerate_available():

import accelerate

from accelerate import dispatch_model

from accelerate.utils import load_offloaded_weights, save_offload_index

def get_parameter_device(parameter: torch.nn.Module) -> torch.device:

from ..hooks.group_offloading import _get_group_onload_device

try:

# Try to get the onload device from the group offloading hook

return _get_group_onload_device(parameter)

except ValueError:

pass

try:

# If the onload device is not available due to no group offloading hooks, try to get the device

# from the first parameter or buffer

parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers())

return next(parameters_and_buffers).device

except StopIteration:

# For torch.nn.DataParallel compatibility in PyTorch 1.5

def find_tensor_attributes(module: torch.nn.Module) -> list[tuple[str, Tensor]]:

tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]

return tuples

gen = parameter._named_members(get_members_fn=find_tensor_attributes)

first_tuple = next(gen)

return first_tuple[1].device

def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:

"""

Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.

"""

# 1. Check if we have attached any dtype modifying hooks (eg. layerwise casting)

if isinstance(parameter, nn.Module):

for name, submodule in parameter.named_modules():

if not hasattr(submodule, "_diffusers_hook"):

continue

registry = submodule._diffusers_hook

hook = registry.get_hook("layerwise_casting")

if hook is not None:

return hook.compute_dtype

# 2. If no dtype modifying hooks are attached, return the dtype of the first floating point parameter/buffer

last_dtype = None

for name, param in parameter.named_parameters():

last_dtype = param.dtype

if (

hasattr(parameter, "_keep_in_fp32_modules")

and parameter._keep_in_fp32_modules

and any(m in name for m in parameter._keep_in_fp32_modules)

):

continue

if param.is_floating_point():

return param.dtype

for buffer in parameter.buffers():

last_dtype = buffer.dtype

if buffer.is_floating_point():

return buffer.dtype

if last_dtype is not None:

# if no floating dtype was found return whatever the first dtype is

return last_dtype

# For nn.DataParallel compatibility in PyTorch > 1.5

def find_tensor_attributes(module: nn.Module) -> list[tuple[str, Tensor]]:

tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]

return tuples

gen = parameter._named_members(get_members_fn=find_tensor_attributes)

last_tuple = None

for tuple in gen:

last_tuple = tuple

if tuple[1].is_floating_point():

return tuple[1].dtype

if last_tuple is not None:

# fallback to the last dtype

return last_tuple[1].dtype

@contextmanager

def no_init_weights():

"""

Context manager to globally disable weight initialization to speed up loading large models. To do that, all the

torch.nn.init function are all replaced with skip.

"""

def _skip_init(*args, **kwargs):

pass

for name, init_func in TORCH_INIT_FUNCTIONS.items():

setattr(torch.nn.init, name, _skip_init)

try:

yield

finally:

# Restore the original initialization functions

for name, init_func in TORCH_INIT_FUNCTIONS.items():

setattr(torch.nn.init, name, init_func)

class ModelMixin(torch.nn.Module, PushToHubMixin):

r"""

Base class for all models.

[`ModelMixin`] takes care of storing the model configuration and provides methods for loading, downloading and

saving models.

- **config_name** ([`str`]) -- Filename to save a model to when calling [`~models.ModelMixin.save_pretrained`].

"""

config_name = CONFIG_NAME

_automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]

_supports_gradient_checkpointing = False

_keys_to_ignore_on_load_unexpected = None

_no_split_modules = None

_keep_in_fp32_modules = None

_skip_layerwise_casting_patterns = None

_supports_group_offloading = True

_repeated_blocks = []

_parallel_config = None

_cp_plan = None

_skip_keys = None

def __init__(self):

super().__init__()

self._gradient_checkpointing_func = None

def __getattr__(self, name: str) -> Any:

"""The only reason we overwrite `getattr` here is to gracefully deprecate accessing

config attributes directly. See https://github.com/huggingface/diffusers/pull/3129 We need to overwrite

__getattr__ here in addition so that we don't trigger `torch.nn.Module`'s __getattr__':

https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module

"""

is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name)

is_attribute = name in self.__dict__

if is_in_config and not is_attribute:

deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'unet.config.{name}'."

deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False, stacklevel=3)

return self._internal_dict[name]

# call PyTorch's https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module

return super().__getattr__(name)

@property

def is_gradient_checkpointing(self) -> bool:

"""

Whether gradient checkpointing is activated for this model or not.

"""

return any(hasattr(m, "gradient_checkpointing") and m.gradient_checkpointing for m in self.modules())

def enable_gradient_checkpointing(self, gradient_checkpointing_func: Callable | None = None) -> None:

"""

Activates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or

*checkpoint activations* in other frameworks).

Args:

gradient_checkpointing_func (`Callable`, *optional*):

The function to use for gradient checkpointing. If `None`, the default PyTorch checkpointing function

is used (`torch.utils.checkpoint.checkpoint`).

"""

if not self._supports_gradient_checkpointing:

raise ValueError(

f"{self.__class__.__name__} does not support gradient checkpointing. Please make sure to set the boolean attribute "

f"`_supports_gradient_checkpointing` to `True` in the class definition."

)

if gradient_checkpointing_func is None:

def _gradient_checkpointing_func(module, *args):

ckpt_kwargs = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}

return torch.utils.checkpoint.checkpoint(

module.__call__,

*args,

**ckpt_kwargs,

)

gradient_checkpointing_func = _gradient_checkpointing_func

self._set_gradient_checkpointing(enable=True, gradient_checkpointing_func=gradient_checkpointing_func)

def disable_gradient_checkpointing(self) -> None:

"""

Deactivates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or

*checkpoint activations* in other frameworks).

"""

if self._supports_gradient_checkpointing:

self._set_gradient_checkpointing(enable=False)

def set_use_npu_flash_attention(self, valid: bool) -> None:

r"""

Set the switch for the npu flash attention.

"""

def fn_recursive_set_npu_flash_attention(module: torch.nn.Module):

if hasattr(module, "set_use_npu_flash_attention"):

module.set_use_npu_flash_attention(valid)

for child in module.children():

fn_recursive_set_npu_flash_attention(child)

for module in self.children():

if isinstance(module, torch.nn.Module):

fn_recursive_set_npu_flash_attention(module)

def enable_npu_flash_attention(self) -> None:

r"""

Enable npu flash attention from torch_npu

"""

self.set_use_npu_flash_attention(True)

def disable_npu_flash_attention(self) -> None:

r"""

disable npu flash attention from torch_npu

"""

self.set_use_npu_flash_attention(False)

def set_use_xla_flash_attention(

self, use_xla_flash_attention: bool, partition_spec: Callable | None = None, **kwargs

) -> None:

# Recursively walk through all the children.

# Any children which exposes the set_use_xla_flash_attention method

# gets the message

def fn_recursive_set_flash_attention(module: torch.nn.Module):

if hasattr(module, "set_use_xla_flash_attention"):

module.set_use_xla_flash_attention(use_xla_flash_attention, partition_spec, **kwargs)

for child in module.children():

fn_recursive_set_flash_attention(child)

for module in self.children():

if isinstance(module, torch.nn.Module):

fn_recursive_set_flash_attention(module)

def enable_xla_flash_attention(self, partition_spec: Callable | None = None, **kwargs):

r"""

Enable the flash attention pallals kernel for torch_xla.

"""

self.set_use_xla_flash_attention(True, partition_spec, **kwargs)

def disable_xla_flash_attention(self):

r"""

Disable the flash attention pallals kernel for torch_xla.

"""

self.set_use_xla_flash_attention(False)

def set_use_memory_efficient_attention_xformers(self, valid: bool, attention_op: Callable | None = 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)

for module in self.children():

if isinstance(module, torch.nn.Module):

fn_recursive_set_mem_eff(module)

def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None) -> 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.

> [!WARNING] > ⚠️ 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 UNet2DConditionModel

>>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp

>>> model = UNet2DConditionModel.from_pretrained(

... "stabilityai/stable-diffusion-2-1", subfolder="unet", torch_dtype=torch.float16

... )

>>> model = model.to("cuda")

>>> model.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)

```

"""

self.set_use_memory_efficient_attention_xformers(True, attention_op)

def disable_xformers_memory_efficient_attention(self) -> None:

r"""

Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).

"""

self.set_use_memory_efficient_attention_xformers(False)

def enable_layerwise_casting(

self,

storage_dtype: torch.dtype = torch.float8_e4m3fn,

compute_dtype: torch.dtype | None = None,

skip_modules_pattern: tuple[str, ...] | None = None,

skip_modules_classes: tuple[Type[torch.nn.Module], ...] | None = None,

non_blocking: bool = False,

) -> None:

r"""

Activates layerwise casting for the current model.

Layerwise casting is a technique that casts the model weights to a lower precision dtype for storage but

upcasts them on-the-fly to a higher precision dtype for computation. This process can significantly reduce the

memory footprint from model weights, but may lead to some quality degradation in the outputs. Most degradations

are negligible, mostly stemming from weight casting in normalization and modulation layers.

By default, most models in diffusers set the `_skip_layerwise_casting_patterns` attribute to ignore patch

embedding, positional embedding and normalization layers. This is because these layers are most likely

precision-critical for quality. If you wish to change this behavior, you can set the

`_skip_layerwise_casting_patterns` attribute to `None`, or call

[`~hooks.layerwise_casting.apply_layerwise_casting`] with custom arguments.

Example:

Using [`~models.ModelMixin.enable_layerwise_casting`]:

```python

>>> from diffusers import CogVideoXTransformer3DModel

>>> transformer = CogVideoXTransformer3DModel.from_pretrained(

... "THUDM/CogVideoX-5b", subfolder="transformer", torch_dtype=torch.bfloat16

... )

>>> # Enable layerwise casting via the model, which ignores certain modules by default

>>> transformer.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)

```

Args:

storage_dtype (`torch.dtype`):

The dtype to which the model should be cast for storage.

compute_dtype (`torch.dtype`):

The dtype to which the model weights should be cast during the forward pass.

skip_modules_pattern (`tuple[str, ...]`, *optional*):

A list of patterns to match the names of the modules to skip during the layerwise casting process. If

set to `None`, default skip patterns are used to ignore certain internal layers of modules and PEFT

layers.

skip_modules_classes (`tuple[Type[torch.nn.Module], ...]`, *optional*):

A list of module classes to skip during the layerwise casting process.

non_blocking (`bool`, *optional*, defaults to `False`):

If `True`, the weight casting operations are non-blocking.

"""

from ..hooks import apply_layerwise_casting

user_provided_patterns = True

if skip_modules_pattern is None:

from ..hooks.layerwise_casting import DEFAULT_SKIP_MODULES_PATTERN

skip_modules_pattern = DEFAULT_SKIP_MODULES_PATTERN

user_provided_patterns = False

if self._keep_in_fp32_modules is not None:

skip_modules_pattern += tuple(self._keep_in_fp32_modules)

if self._skip_layerwise_casting_patterns is not None:

skip_modules_pattern += tuple(self._skip_layerwise_casting_patterns)

skip_modules_pattern = tuple(set(skip_modules_pattern))

if is_peft_available() and not user_provided_patterns:

# By default, we want to skip all peft layers because they have a very low memory footprint.

# If users want to apply layerwise casting on peft layers as well, they can utilize the

# `~diffusers.hooks.layerwise_casting.apply_layerwise_casting` function which provides

# them with more flexibility and control.

from peft.tuners.loha.layer import LoHaLayer

from peft.tuners.lokr.layer import LoKrLayer

from peft.tuners.lora.layer import LoraLayer

for layer in (LoHaLayer, LoKrLayer, LoraLayer):

skip_modules_pattern += tuple(layer.adapter_layer_names)

if compute_dtype is None:

logger.info("`compute_dtype` not provided when enabling layerwise casting. Using dtype of the model.")

compute_dtype = self.dtype

apply_layerwise_casting(

self, storage_dtype, compute_dtype, skip_modules_pattern, skip_modules_classes, non_blocking

)

def enable_group_offload(

self,

onload_device: torch.device,

offload_device: torch.device = torch.device("cpu"),

offload_type: str = "block_level",

num_blocks_per_group: int | None = None,

non_blocking: bool = False,

use_stream: bool = False,

record_stream: bool = False,

low_cpu_mem_usage=False,

offload_to_disk_path: str | None = None,

block_modules: str | None = None,

exclude_kwargs: str | None = None,

) -> None:

r"""

Activates group offloading for the current model.

See [`~hooks.group_offloading.apply_group_offloading`] for more information.

Example:

```python

>>> from diffusers import CogVideoXTransformer3DModel

>>> transformer = CogVideoXTransformer3DModel.from_pretrained(

... "THUDM/CogVideoX-5b", subfolder="transformer", torch_dtype=torch.bfloat16

... )

>>> transformer.enable_group_offload(

... onload_device=torch.device("cuda"),

... offload_device=torch.device("cpu"),

... offload_type="leaf_level",

... use_stream=True,

... )

```

"""

from ..hooks import apply_group_offloading

if getattr(self, "enable_tiling", None) is not None and getattr(self, "use_tiling", False) and use_stream:

msg = (

"Applying group offloading on autoencoders, with CUDA streams, may not work as expected if the first "

"forward pass is executed with tiling enabled. Please make sure to either:\n"

"1. Run a forward pass with small input shapes.\n"

"2. Or, run a forward pass with tiling disabled (can still use small dummy inputs)."

)

logger.warning(msg)

if not self._supports_group_offloading:

raise ValueError(

f"{self.__class__.__name__} does not support group offloading. Please make sure to set the boolean attribute "

f"`_supports_group_offloading` to `True` in the class definition. If you believe this is a mistake, please "

f"open an issue at https://github.com/huggingface/diffusers/issues."

)

apply_group_offloading(

module=self,

onload_device=onload_device,

offload_device=offload_device,

offload_type=offload_type,

num_blocks_per_group=num_blocks_per_group,

non_blocking=non_blocking,

use_stream=use_stream,

record_stream=record_stream,

low_cpu_mem_usage=low_cpu_mem_usage,

offload_to_disk_path=offload_to_disk_path,

block_modules=block_modules,

exclude_kwargs=exclude_kwargs,

)

def set_attention_backend(self, backend: str) -> None:

"""

Set the attention backend for the model.

Args:

backend (`str`):

The name of the backend to set. Must be one of the available backends defined in

`AttentionBackendName`. Available backends can be found in

`diffusers.attention_dispatch.AttentionBackendName`. Defaults to torch native scaled dot product

attention as backend.

"""

from .attention import AttentionModuleMixin

from .attention_dispatch import (

AttentionBackendName,

_AttentionBackendRegistry,

_check_attention_backend_requirements,

_maybe_download_kernel_for_backend,

)

# TODO: the following will not be required when everything is refactored to AttentionModuleMixin

from .attention_processor import Attention, MochiAttention

logger.warning("Attention backends are an experimental feature and the API may be subject to change.")

attention_classes = (Attention, MochiAttention, AttentionModuleMixin)

parallel_config_set = False

for module in self.modules():

if not isinstance(module, attention_classes):

continue

processor = module.processor

if getattr(processor, "_parallel_config", None) is not None:

parallel_config_set = True

break

backend = backend.lower()

available_backends = {x.value for x in AttentionBackendName.__members__.values()}

if backend not in available_backends:

raise ValueError(f"`{backend=}` must be one of the following: " + ", ".join(available_backends))

backend = AttentionBackendName(backend)

if parallel_config_set and not _AttentionBackendRegistry._is_context_parallel_available(backend):

compatible_backends = sorted(_AttentionBackendRegistry._supports_context_parallel)

raise ValueError(

f"Context parallelism is enabled but current attention backend '{backend.value}' "

f"does not support context parallelism. "

f"Please set a compatible attention backend: {compatible_backends} using `model.set_attention_backend()`."

)

_check_attention_backend_requirements(backend)

_maybe_download_kernel_for_backend(backend)

for module in self.modules():

if not isinstance(module, attention_classes):

continue

processor = module.processor

if processor is None or not hasattr(processor, "_attention_backend"):

continue

processor._attention_backend = backend

# Important to set the active backend so that it propagates gracefully throughout.

_AttentionBackendRegistry.set_active_backend(backend)

def reset_attention_backend(self) -> None:

"""

Resets the attention backend for the model. Following calls to `forward` will use the environment default, if

set, or the torch native scaled dot product attention.

"""

from .attention import AttentionModuleMixin

from .attention_processor import Attention, MochiAttention

logger.warning("Attention backends are an experimental feature and the API may be subject to change.")

attention_classes = (Attention, MochiAttention, AttentionModuleMixin)

for module in self.modules():

if not isinstance(module, attention_classes):

continue

processor = module.processor

if processor is None or not hasattr(processor, "_attention_backend"):

continue

processor._attention_backend = None

def save_pretrained(

self,

save_directory: str | os.PathLike,

is_main_process: bool = True,

save_function: Callable | None = None,

safe_serialization: bool = True,

variant: str | None = None,

max_shard_size: int | str = "10GB",

push_to_hub: bool = False,

**kwargs,

):

"""

Save a model and its configuration file to a directory so that it can be reloaded using the

[`~models.ModelMixin.from_pretrained`] class method.

Arguments:

save_directory (`str` or `os.PathLike`):

Directory to save a model and its configuration file to. Will be created if it doesn't exist.

is_main_process (`bool`, *optional*, defaults to `True`):

Whether the process calling this is the main process or not. Useful during distributed training and you

need to call this function on all processes. In this case, set `is_main_process=True` only on the main

process to avoid race conditions.

save_function (`Callable`):

The function to use to save the state dictionary. Useful during distributed training when you need to

replace `torch.save` with another method. Can be configured with the environment variable

`DIFFUSERS_SAVE_MODE`.

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.<variant>.bin`.

max_shard_size (`int` or `str`, defaults to `"10GB"`):

The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size

lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5GB"`).

If expressed as an integer, the unit is bytes. Note that this limit will be decreased after a certain

period of time (starting from Oct 2024) to allow users to upgrade to the latest version of `diffusers`.

This is to establish a common default size for this argument across different libraries in the Hugging

Face ecosystem (`transformers`, and `accelerate`, for example).

push_to_hub (`bool`, *optional*, defaults to `False`):

Whether or not to push your model to the Hugging Face 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.

"""

if os.path.isfile(save_directory):

logger.error(f"Provided path ({save_directory}) should be a directory, not a file")

return

hf_quantizer = getattr(self, "hf_quantizer", None)

if hf_quantizer is not None:

quantization_serializable = (

hf_quantizer is not None

and isinstance(hf_quantizer, DiffusersQuantizer)

and hf_quantizer.is_serializable

)

if not quantization_serializable:

raise ValueError(

f"The model is quantized with {hf_quantizer.quantization_config.quant_method} and is not serializable - check out the warnings from"

" the logger on the traceback to understand the reason why the quantized model is not serializable."

)

weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME

weights_name = _add_variant(weights_name, variant)

weights_name_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(

".safetensors", "{suffix}.safetensors"

)

os.makedirs(save_directory, exist_ok=True)

if push_to_hub:

commit_message = kwargs.pop("commit_message", None)

private = kwargs.pop("private", None)

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

# Only save the model itself if we are using distributed training

model_to_save = self

# Attach architecture to the config

# Save the config

if is_main_process:

model_to_save.save_config(save_directory)

# Save the model

state_dict = model_to_save.state_dict()

# Save the model

state_dict_split = split_torch_state_dict_into_shards(

state_dict, max_shard_size=max_shard_size, filename_pattern=weights_name_pattern

)

# Clean the folder from a previous save

if is_main_process:

for filename in os.listdir(save_directory):

if filename in state_dict_split.filename_to_tensors.keys():

continue

full_filename = os.path.join(save_directory, filename)

if not os.path.isfile(full_filename):

continue

weights_without_ext = weights_name_pattern.replace(".bin", "").replace(".safetensors", "")

weights_without_ext = weights_without_ext.replace("{suffix}", "")

filename_without_ext = filename.replace(".bin", "").replace(".safetensors", "")

# make sure that file to be deleted matches format of sharded file, e.g. pytorch_model-00001-of-00005

if (

filename.startswith(weights_without_ext)

and _REGEX_SHARD.fullmatch(filename_without_ext) is not None

):

os.remove(full_filename)

for filename, tensors in state_dict_split.filename_to_tensors.items():

shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}

filepath = os.path.join(save_directory, filename)

if safe_serialization:

# At some point we will need to deal better with save_function (used for TPU and other distributed

# joyfulness), but for now this enough.

safetensors.torch.save_file(shard, filepath, metadata={"format": "pt"})

else:

torch.save(shard, filepath)

if state_dict_split.is_sharded:

index = {

"metadata": state_dict_split.metadata,

"weight_map": state_dict_split.tensor_to_filename,

}

save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME

save_index_file = os.path.join(save_directory, _add_variant(save_index_file, variant))

# Save the index as well

with open(save_index_file, "w", encoding="utf-8") as f:

content = json.dumps(index, indent=2, sort_keys=True) + "\n"

f.write(content)

logger.info(

f"The model is bigger than the maximum size per checkpoint ({max_shard_size}) and is going to be "

f"split in {len(state_dict_split.filename_to_tensors)} checkpoint shards. You can find where each parameters has been saved in the "

f"index located at {save_index_file}."

)

else:

path_to_weights = os.path.join(save_directory, weights_name)

logger.info(f"Model weights saved in {path_to_weights}")

if push_to_hub:

# Create a new empty model card and eventually tag it

model_card = load_or_create_model_card(repo_id, token=token)

model_card = populate_model_card(model_card)

model_card.save(Path(save_directory, "README.md").as_posix())

self._upload_folder(

save_directory,

repo_id,

token=token,

commit_message=commit_message,

create_pr=create_pr,

)

def dequantize(self):

"""

Potentially dequantize the model in case it has been quantized by a quantization method that support

dequantization.

"""

hf_quantizer = getattr(self, "hf_quantizer", None)

if hf_quantizer is None:

raise ValueError("You need to first quantize your model in order to dequantize it")

return hf_quantizer.dequantize(self)

@classmethod

@validate_hf_hub_args

def from_pretrained(cls, pretrained_model_name_or_path: str | os.PathLike | None, **kwargs) -> Self:

r"""

Instantiate a pretrained PyTorch model from a pretrained model configuration.

The model is set in evaluation mode - `model.eval()` - by default, and dropout modules are deactivated. To

train the model, set it back in training mode with `model.train()`.

Parameters:

pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):

Can be either:

- A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on

the Hub.

- A path to a *directory* (for example `./my_model_directory`) containing the model weights saved

with [`~ModelMixin.save_pretrained`].

cache_dir (`str | os.PathLike`, *optional*):

Path to a directory where a downloaded pretrained model configuration is cached if the standard cache

is not used.

torch_dtype (`torch.dtype`, *optional*):

Override the default `torch.dtype` and load the model with another dtype.

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.

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.

from_flax (`bool`, *optional*, defaults to `False`):

Load the model weights from a Flax checkpoint save file.

subfolder (`str`, *optional*, defaults to `""`):

The subfolder location of a model file within a larger model repository on the Hub or locally.

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 (`int | str | torch.device` or `dict[str, 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. Defaults to `None`, meaning that the model will be loaded on CPU.

Examples:

```py

>>> from diffusers import AutoModel

>>> import torch

>>> # This works.

>>> model = AutoModel.from_pretrained(

... "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", device_map="cuda"

... )

>>> # This also works (integer accelerator device ID).

>>> model = AutoModel.from_pretrained(

... "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", device_map=0

... )

>>> # Specifying a supported offloading strategy like "auto" also works.

>>> model = AutoModel.from_pretrained(

... "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", device_map="auto"

... )

>>> # Specifying a dictionary as `device_map` also works.

>>> model = AutoModel.from_pretrained(

... "stabilityai/stable-diffusion-xl-base-1.0",

... subfolder="unet",

... device_map={"": torch.device("cuda")},

... )

```

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://huggingface.co/docs/accelerate/en/concept_guides/big_model_inference#the-devicemap). You

can also refer to the [Diffusers-specific

documentation](https://huggingface.co/docs/diffusers/main/en/training/distributed_inference#model-sharding)

for more concrete examples.

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.

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.

disable_mmap ('bool', *optional*, defaults to 'False'):

Whether to disable mmap when loading a Safetensors model. This option can perform better when the model

is on a network mount or hard drive, which may not handle the seeky-ness of mmap very well.

> [!TIP] > To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in

with `hf > auth login`. You can also activate the special >

["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a >

firewalled environment.

Example:

```py

from diffusers import UNet2DConditionModel

unet = UNet2DConditionModel.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", subfolder="unet")

```

If you get the error message below, you need to finetune the weights for your downstream task:

```bash

Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/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.

```

"""

cache_dir = kwargs.pop("cache_dir", None)

ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)

force_download = kwargs.pop("force_download", False)

from_flax = kwargs.pop("from_flax", False)

proxies = kwargs.pop("proxies", None)

output_loading_info = kwargs.pop("output_loading_info", False)

local_files_only = kwargs.pop("local_files_only", None)

token = kwargs.pop("token", None)

revision = kwargs.pop("revision", None)

torch_dtype = kwargs.pop("torch_dtype", None)

subfolder = kwargs.pop("subfolder", 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", None)

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)

quantization_config = kwargs.pop("quantization_config", None)

dduf_entries: dict[str, DDUFEntry] | None = kwargs.pop("dduf_entries", None)

disable_mmap = kwargs.pop("disable_mmap", False)

parallel_config: ParallelConfig | ContextParallelConfig | None = kwargs.pop("parallel_config", None)

is_parallel_loading_enabled = HF_ENABLE_PARALLEL_LOADING

if is_parallel_loading_enabled and not low_cpu_mem_usage:

raise NotImplementedError("Parallel loading is not supported when not using `low_cpu_mem_usage`.")

if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):

torch_dtype = torch.float32

logger.warning(

f"Passed `torch_dtype` {torch_dtype} is not a `torch.dtype`. Defaulting to `torch.float32`."

)

allow_pickle = False

if use_safetensors is None:

use_safetensors = True