Models (original) (raw)

🤗 Diffusers provides pretrained models for popular algorithms and modules to create custom diffusion systems. The primary function of models is to denoise an input sample as modeled by the distribution pθ(xt−1∣xt)p_{\theta}(x_{t-1}|x_{t}).

All models are built from the base ModelMixin class which is a torch.nn.Module providing basic functionality for saving and loading models, locally and from the Hugging Face Hub.

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 save_pretrained().

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

Deactivates gradient checkpointing for the current model (may be referred to as activation checkpointing or_checkpoint activations_ in other frameworks).

disable npu flash attention from torch_npu

disable_xformers_memory_efficient_attention

( )

Disable memory efficient attention from xFormers.

Disable the flash attention pallals kernel for torch_xla.

enable_gradient_checkpointing

( gradient_checkpointing_func: typing.Optional[typing.Callable] = None )

Parameters

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).

Activates gradient checkpointing for the current model (may be referred to as activation checkpointing or_checkpoint activations_ in other frameworks).

enable_group_offload

( onload_device: device offload_device: device = device(type='cpu') offload_type: str = 'block_level' num_blocks_per_group: typing.Optional[int] = None non_blocking: bool = False use_stream: bool = False record_stream: bool = False low_cpu_mem_usage = False )

Activates group offloading for the current model.

See apply_group_offloading() for more information.

Example:

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, ... )

enable_layerwise_casting

( storage_dtype: dtype = torch.float8_e4m3fn compute_dtype: typing.Optional[torch.dtype] = None skip_modules_pattern: typing.Optional[typing.Tuple[str, ...]] = None skip_modules_classes: typing.Optional[typing.Tuple[typing.Type[torch.nn.modules.module.Module], ...]] = None non_blocking: bool = False )

Parameters

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.

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 callapply_layerwise_casting() with custom arguments.

Example:

Using enable_layerwise_casting():

from diffusers import CogVideoXTransformer3DModel

transformer = CogVideoXTransformer3DModel.from_pretrained( ... "THUDM/CogVideoX-5b", subfolder="transformer", torch_dtype=torch.bfloat16 ... )

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

Enable npu flash attention from torch_npu

enable_xformers_memory_efficient_attention

( attention_op: typing.Optional[typing.Callable] = None )

Parameters

attention_op (Callable, optional) — Override the default None operator for use as op argument to thememory_efficient_attention()function of xFormers.

Enable memory efficient attention from 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.

Examples:

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)

enable_xla_flash_attention

( partition_spec: typing.Optional[typing.Callable] = None **kwargs )

Enable the flash attention pallals kernel for torch_xla.

from_pretrained

( pretrained_model_name_or_path: typing.Union[str, os.PathLike, NoneType] **kwargs )

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 save_pretrained().
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.
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.
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 fromdiffusers-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 (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. Defaults to None, meaning that the model will be loaded on CPU.
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.
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 Truewhen 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 thesafetensors library is installed. If set to True, the model is forcibly loaded from safetensorsweights. 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.

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().

To use private or gated models, log-in withhuggingface-cli login. You can also activate the special“offline-mode” to use this method in a firewalled environment.

Example:

from diffusers import UNet2DConditionModel

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

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.

get_memory_footprint

( return_buffers = True )

Parameters

return_buffers (bool, optional, defaults to True) — Whether to return the size of the buffer tensors in the computation of the memory footprint. Buffers are tensors that do not require gradients and not registered as parameters. E.g. mean and std in batch norm layers. Please see: https://discuss.pytorch.org/t/what-pytorch-means-by-buffers/120266/2

Get the memory footprint of a model. This will return the memory footprint of the current model in bytes. Useful to benchmark the memory footprint of the current model and design some tests. Solution inspired from the PyTorch discussions: https://discuss.pytorch.org/t/gpu-memory-that-model-uses/56822/2

num_parameters

( only_trainable: bool = False exclude_embeddings: bool = False ) → int

Parameters

only_trainable (bool, optional, defaults to False) — Whether or not to return only the number of trainable parameters.
exclude_embeddings (bool, optional, defaults to False) — Whether or not to return only the number of non-embedding parameters.

The number of parameters.

Get number of (trainable or non-embedding) parameters in the module.

Example:

from diffusers import UNet2DConditionModel

model_id = "runwayml/stable-diffusion-v1-5" unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet") unet.num_parameters(only_trainable=True) 859520964

save_pretrained

( save_directory: typing.Union[str, os.PathLike] is_main_process: bool = True save_function: typing.Optional[typing.Callable] = None safe_serialization: bool = True variant: typing.Optional[str] = None max_shard_size: typing.Union[int, str] = '10GB' push_to_hub: bool = False **kwargs )

Parameters

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 variableDIFFUSERS_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 push_to_hub() method.

Save a model and its configuration file to a directory so that it can be reloaded using thefrom_pretrained() class method.

set_use_npu_flash_attention

( valid: bool )

Set the switch for the npu flash attention.

Base class for all Flax models.

FlaxModelMixin 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 save_pretrained().

from_pretrained

( pretrained_model_name_or_path: typing.Union[str, os.PathLike] dtype: dtype = <class 'jax.numpy.float32'> *model_args **kwargs )

Parameters

pretrained_model_name_or_path (str or os.PathLike) — Can be either:
- A string, the model id (for example runwayml/stable-diffusion-v1-5) of a pretrained model hosted on the Hub.
- A path to a directory (for example ./my_model_directory) containing the model weights saved using save_pretrained().
dtype (jax.numpy.dtype, optional, defaults to jax.numpy.float32) — The data type of the computation. Can be one of jax.numpy.float32, jax.numpy.float16 (on GPUs) andjax.numpy.bfloat16 (on TPUs).
This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If specified, all the computation will be performed with the given dtype.
This only specifies the dtype of the computation and does not influence the dtype of model parameters.
If you wish to change the dtype of the model parameters, see to_fp16() andto_bf16().
model_args (sequence of positional arguments, optional) — All remaining positional arguments are passed to the underlying model’s __init__ method.
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.
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.
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.
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_pt (bool, optional, defaults to False) — Load the model weights from a PyTorch checkpoint save file.
kwargs (remaining dictionary of keyword arguments, optional) — Can be used to update the configuration object (after it is loaded) and initiate the model (for example, output_attentions=True). Behaves differently depending on whether a config is provided or automatically loaded:
- If a configuration is provided with config, kwargs are directly passed to the underlying model’s __init__ method (we assume all relevant updates to the configuration have already been done).
- If a configuration is not provided, kwargs are first passed to the configuration class initialization function from_config(). Each key of the kwargs that corresponds to a configuration attribute is used to override said attribute with the supplied kwargs value. Remaining keys that do not correspond to any configuration attribute are passed to the underlying model’s __init__ function.

Instantiate a pretrained Flax model from a pretrained model configuration.

Examples:

from diffusers import FlaxUNet2DConditionModel

model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")

model, params = FlaxUNet2DConditionModel.from_pretrained("./test/saved_model/")

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.

save_pretrained

( save_directory: typing.Union[str, os.PathLike] params: typing.Union[typing.Dict, flax.core.frozen_dict.FrozenDict] is_main_process: bool = True push_to_hub: bool = False **kwargs )

Parameters

save_directory (str or os.PathLike) — Directory to save a model and its configuration file to. Will be created if it doesn’t exist.
params (Union[Dict, FrozenDict]) — A PyTree of model parameters.
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.
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 key word arguments passed along to the push_to_hub() method.

Save a model and its configuration file to a directory so that it can be reloaded using thefrom_pretrained() class method.

to_bf16

( params: typing.Union[typing.Dict, flax.core.frozen_dict.FrozenDict] mask: typing.Any = None )

Parameters

params (Union[Dict, FrozenDict]) — A PyTree of model parameters.
mask (Union[Dict, FrozenDict]) — A PyTree with same structure as the params tree. The leaves should be booleans. It should be Truefor params you want to cast, and False for those you want to skip.

Cast the floating-point params to jax.numpy.bfloat16. This returns a new params tree and does not cast the params in place.

This method can be used on a TPU to explicitly convert the model parameters to bfloat16 precision to do full half-precision training or to save weights in bfloat16 for inference in order to save memory and improve speed.

Examples:

from diffusers import FlaxUNet2DConditionModel

model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")

params = model.to_bf16(params)

from flax import traverse_util

model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5") flat_params = traverse_util.flatten_dict(params) mask = { ... path: (path[-2] != ("LayerNorm", "bias") and path[-2:] != ("LayerNorm", "scale")) ... for path in flat_params ... } mask = traverse_util.unflatten_dict(mask) params = model.to_bf16(params, mask)

to_fp16

( params: typing.Union[typing.Dict, flax.core.frozen_dict.FrozenDict] mask: typing.Any = None )

Parameters

params (Union[Dict, FrozenDict]) — A PyTree of model parameters.
mask (Union[Dict, FrozenDict]) — A PyTree with same structure as the params tree. The leaves should be booleans. It should be Truefor params you want to cast, and False for those you want to skip.

Cast the floating-point params to jax.numpy.float16. This returns a new params tree and does not cast theparams in place.

This method can be used on a GPU to explicitly convert the model parameters to float16 precision to do full half-precision training or to save weights in float16 for inference in order to save memory and improve speed.

Examples:

from diffusers import FlaxUNet2DConditionModel

model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")

params = model.to_fp16(params)

from flax import traverse_util

model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5") flat_params = traverse_util.flatten_dict(params) mask = { ... path: (path[-2] != ("LayerNorm", "bias") and path[-2:] != ("LayerNorm", "scale")) ... for path in flat_params ... } mask = traverse_util.unflatten_dict(mask) params = model.to_fp16(params, mask)

to_fp32

( params: typing.Union[typing.Dict, flax.core.frozen_dict.FrozenDict] mask: typing.Any = None )

Parameters

params (Union[Dict, FrozenDict]) — A PyTree of model parameters.
mask (Union[Dict, FrozenDict]) — A PyTree with same structure as the params tree. The leaves should be booleans. It should be Truefor params you want to cast, and False for those you want to skip.

Cast the floating-point params to jax.numpy.float32. This method can be used to explicitly convert the model parameters to fp32 precision. This returns a new params tree and does not cast the params in place.

Examples:

from diffusers import FlaxUNet2DConditionModel

model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")

params = model.to_f16(params)

params = model.to_fp32(params)

class diffusers.utils.PushToHubMixin

( )

A Mixin to push a model, scheduler, or pipeline to the Hugging Face Hub.

push_to_hub

( repo_id: str commit_message: typing.Optional[str] = None private: typing.Optional[bool] = None token: typing.Optional[str] = None create_pr: bool = False safe_serialization: bool = True variant: typing.Optional[str] = None )

Parameters

repo_id (str) — The name of the repository you want to push your model, scheduler, or pipeline files to. It should contain your organization name when pushing to an organization. repo_id can also be a path to a local directory.
commit_message (str, optional) — Message to commit while pushing. Default to "Upload {object}".
private (bool, optional) — Whether to make the repo private. If None (default), the repo will be public unless the organization’s default is private. This value is ignored if the repo already exists.
token (str, optional) — The token to use as HTTP bearer authorization for remote files. The token generated when runninghuggingface-cli login (stored in ~/.huggingface).
create_pr (bool, optional, defaults to False) — Whether or not to create a PR with the uploaded files or directly commit.
safe_serialization (bool, optional, defaults to True) — Whether or not to convert the model weights to the safetensors format.
variant (str, optional) — If specified, weights are saved in the format pytorch_model.<variant>.bin.

Upload model, scheduler, or pipeline files to the 🤗 Hugging Face Hub.

Examples:

from diffusers import UNet2DConditionModel

unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="unet")

unet.push_to_hub("my-finetuned-unet")

unet.push_to_hub("your-org/my-finetuned-unet")