diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index 34cc16591d40..3804d2e53f5f 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -63,6 +63,7 @@ from .utils.dummy_torch_and_transformers_and_onnx_objects import * # noqa F403 if is_flax_available(): + from .modeling_flax_utils import FlaxModelMixin from .schedulers import FlaxPNDMScheduler else: from .utils.dummy_flax_objects import * # noqa F403 diff --git a/src/diffusers/configuration_utils.py b/src/diffusers/configuration_utils.py index 1294710474f8..cf85eccb44df 100644 --- a/src/diffusers/configuration_utils.py +++ b/src/diffusers/configuration_utils.py @@ -14,6 +14,7 @@ # See the License for the specific language governing permissions and # limitations under the License. """ ConfigMixinuration base class and utilities.""" +import dataclasses import functools import inspect import json @@ -271,6 +272,11 @@ def extract_init_dict(cls, config_dict, **kwargs): # remove general kwargs if present in dict if "kwargs" in expected_keys: expected_keys.remove("kwargs") + # remove flax interal keys + if hasattr(cls, "_flax_internal_args"): + for arg in cls._flax_internal_args: + expected_keys.remove(arg) + # remove keys to be ignored if len(cls.ignore_for_config) > 0: expected_keys = expected_keys - set(cls.ignore_for_config) @@ -401,3 +407,44 @@ def inner_init(self, *args, **kwargs): getattr(self, "register_to_config")(**new_kwargs) return inner_init + + +def flax_register_to_config(cls): + original_init = cls.__init__ + + @functools.wraps(original_init) + def init(self, *args, **kwargs): + if not isinstance(self, ConfigMixin): + raise RuntimeError( + f"`@register_for_config` was applied to {self.__class__.__name__} init method, but this class does " + "not inherit from `ConfigMixin`." + ) + + # Ignore private kwargs in the init. Retrieve all passed attributes + init_kwargs = {k: v for k, v in kwargs.items() if not k.startswith("_")} + + # Retrieve default values + fields = dataclasses.fields(self) + default_kwargs = {} + for field in fields: + # ignore flax specific attributes + if field.name in self._flax_internal_args: + continue + if type(field.default) == dataclasses._MISSING_TYPE: + default_kwargs[field.name] = None + else: + default_kwargs[field.name] = getattr(self, field.name) + + # Make sure init_kwargs override default kwargs + new_kwargs = {**default_kwargs, **init_kwargs} + + # Get positional arguments aligned with kwargs + for i, arg in enumerate(args): + name = fields[i].name + new_kwargs[name] = arg + + getattr(self, "register_to_config")(**new_kwargs) + original_init(self, *args, **kwargs) + + cls.__init__ = init + return cls diff --git a/src/diffusers/modeling_flax_utils.py b/src/diffusers/modeling_flax_utils.py new file mode 100644 index 000000000000..4f2d25dfb168 --- /dev/null +++ b/src/diffusers/modeling_flax_utils.py @@ -0,0 +1,461 @@ +# coding=utf-8 +# Copyright 2022 The HuggingFace Inc. team. +# +# 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 os +from pickle import UnpicklingError +from typing import Any, Dict, Union + +import jax +import jax.numpy as jnp +import msgpack.exceptions +from flax.core.frozen_dict import FrozenDict +from flax.serialization import from_bytes, to_bytes +from flax.traverse_util import flatten_dict, unflatten_dict +from huggingface_hub import hf_hub_download +from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError +from requests import HTTPError + +from .modeling_utils import WEIGHTS_NAME +from .utils import CONFIG_NAME, DIFFUSERS_CACHE, HUGGINGFACE_CO_RESOLVE_ENDPOINT, logging + + +FLAX_WEIGHTS_NAME = "diffusion_flax_model.msgpack" + +logger = logging.get_logger(__name__) + + +class FlaxModelMixin: + r""" + Base class for all flax models. + + [`FlaxModelMixin`] takes care of storing the configuration of the models and handles methods for loading, + downloading and saving models. + """ + config_name = CONFIG_NAME + _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"] + _flax_internal_args = ["name", "parent"] + + @classmethod + def _from_config(cls, config, **kwargs): + """ + All context managers that the model should be initialized under go here. + """ + return cls(config, **kwargs) + + def _cast_floating_to(self, params: Union[Dict, FrozenDict], dtype: jnp.dtype, mask: Any = None) -> Any: + """ + Helper method to cast floating-point values of given parameter `PyTree` to given `dtype`. + """ + + # taken from https://github.com/deepmind/jmp/blob/3a8318abc3292be38582794dbf7b094e6583b192/jmp/_src/policy.py#L27 + def conditional_cast(param): + if isinstance(param, jnp.ndarray) and jnp.issubdtype(param.dtype, jnp.floating): + param = param.astype(dtype) + return param + + if mask is None: + return jax.tree_map(conditional_cast, params) + + flat_params = flatten_dict(params) + flat_mask, _ = jax.tree_flatten(mask) + + for masked, key in zip(flat_mask, flat_params.keys()): + if masked: + param = flat_params[key] + flat_params[key] = conditional_cast(param) + + return unflatten_dict(flat_params) + + def to_bf16(self, params: Union[Dict, FrozenDict], mask: Any = None): + r""" + 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 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. + + Arguments: + 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, `True` for params + you want to cast, and should be `False` for those you want to skip. + + Examples: + + ```python + >>> from diffusers import FlaxUNet2DConditionModel + + >>> # load model + >>> model, params = FlaxUNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4") + >>> # By default, the model parameters will be in fp32 precision, to cast these to bfloat16 precision + >>> params = model.to_bf16(params) + >>> # If you don't want to cast certain parameters (for example layer norm bias and scale) + >>> # then pass the mask as follows + >>> from flax import traverse_util + + >>> model, params = FlaxUNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4") + >>> 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) + ```""" + return self._cast_floating_to(params, jnp.bfloat16, mask) + + def to_fp32(self, params: Union[Dict, FrozenDict], mask: Any = None): + r""" + 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. + + Arguments: + 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, `True` for params + you want to cast, and should be `False` for those you want to skip + + Examples: + + ```python + >>> from diffusers import FlaxUNet2DConditionModel + + >>> # Download model and configuration from huggingface.co + >>> model, params = FlaxUNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4") + >>> # By default, the model params will be in fp32, to illustrate the use of this method, + >>> # we'll first cast to fp16 and back to fp32 + >>> params = model.to_f16(params) + >>> # now cast back to fp32 + >>> params = model.to_fp32(params) + ```""" + return self._cast_floating_to(params, jnp.float32, mask) + + def to_fp16(self, params: Union[Dict, FrozenDict], mask: Any = None): + r""" + Cast the floating-point `params` to `jax.numpy.float16`. This returns a new `params` tree and does not cast the + `params` in place. + + This method can be used on 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. + + Arguments: + 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, `True` for params + you want to cast, and should be `False` for those you want to skip + + Examples: + + ```python + >>> from diffusers import FlaxUNet2DConditionModel + + >>> # load model + >>> model, params = FlaxUNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4") + >>> # By default, the model params will be in fp32, to cast these to float16 + >>> params = model.to_fp16(params) + >>> # If you want don't want to cast certain parameters (for example layer norm bias and scale) + >>> # then pass the mask as follows + >>> from flax import traverse_util + + >>> model, params = FlaxUNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4") + >>> 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) + ```""" + return self._cast_floating_to(params, jnp.float16, mask) + + @classmethod + def from_pretrained( + cls, + pretrained_model_name_or_path: Union[str, os.PathLike], + dtype: jnp.dtype = jnp.float32, + *model_args, + **kwargs, + ): + r""" + Instantiate a pretrained flax model from a pre-trained model configuration. + + The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come + pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning + task. + + The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those + weights are discarded. + + Parameters: + pretrained_model_name_or_path (`str` or `os.PathLike`): + Can be either: + + - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co. + Valid model ids are namespaced under a user or organization name, like + `CompVis/stable-diffusion-v1-4`. + - A path to a *directory* containing model weights saved using [`~ModelMixin.save_pretrained`], + e.g., `./my_model_directory/`. + 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) and + `jax.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`. + + **Note that 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 [`~ModelMixin.to_fp16`] and + [`~ModelMixin.to_bf16`]. + model_args (sequence of positional arguments, *optional*): + All remaining positional arguments will be passed to the underlying model's `__init__` method. + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory in which a downloaded pretrained model configuration should be cached if the + standard cache should not be used. + ignore_mismatched_sizes (`bool`, *optional*, defaults to `False`): + Whether or not to raise an error if some of the weights from the checkpoint do not have the same size + as the weights of the model (if for instance, you are instantiating a model with 10 labels from a + checkpoint with 3 labels). + 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 delete incompletely received files. Will attempt to resume the download if such a + file exists. + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'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 or not to only look at local files (i.e., do not try to download the model). + 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, since we use a + git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any + identifier allowed by git. + kwargs (remaining dictionary of keyword arguments, *optional*): + Can be used to update the configuration object (after it being loaded) and initiate the model (e.g., + `output_attentions=True`). Behaves differently depending on whether a `config` is provided or + automatically loaded: + + - If a configuration is provided with `config`, `**kwargs` will be 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` will be first passed to the configuration class + initialization function ([`~ConfigMixin.from_config`]). Each key of `kwargs` that corresponds to + a configuration attribute will be used to override said attribute with the supplied `kwargs` + value. Remaining keys that do not correspond to any configuration attribute will be passed to the + underlying model's `__init__` function. + + Examples: + + ```python + >>> from diffusers import FlaxUNet2DConditionModel + + >>> # Download model and configuration from huggingface.co and cache. + >>> model, params = FlaxUNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4") + >>> # Model was saved using *save_pretrained('./test/saved_model/')* (for example purposes, not runnable). + >>> model, params = FlaxUNet2DConditionModel.from_pretrained("./test/saved_model/") + ```""" + config = kwargs.pop("config", None) + cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE) + force_download = kwargs.pop("force_download", False) + resume_download = kwargs.pop("resume_download", False) + proxies = kwargs.pop("proxies", None) + local_files_only = kwargs.pop("local_files_only", False) + use_auth_token = kwargs.pop("use_auth_token", None) + revision = kwargs.pop("revision", None) + from_auto_class = kwargs.pop("_from_auto", False) + subfolder = kwargs.pop("subfolder", None) + + user_agent = {"file_type": "model", "framework": "flax", "from_auto_class": from_auto_class} + + # Load config if we don't provide a configuration + config_path = config if config is not None else pretrained_model_name_or_path + model, model_kwargs = cls.from_config( + config_path, + cache_dir=cache_dir, + return_unused_kwargs=True, + force_download=force_download, + resume_download=resume_download, + proxies=proxies, + local_files_only=local_files_only, + use_auth_token=use_auth_token, + revision=revision, + # model args + dtype=dtype, + **kwargs, + ) + + # Load model + if os.path.isdir(pretrained_model_name_or_path): + if os.path.isfile(os.path.join(pretrained_model_name_or_path, FLAX_WEIGHTS_NAME)): + # Load from a Flax checkpoint + model_file = os.path.join(pretrained_model_name_or_path, FLAX_WEIGHTS_NAME) + # At this stage we don't have a weight file so we will raise an error. + elif os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME): + raise EnvironmentError( + f"Error no file named {FLAX_WEIGHTS_NAME} found in directory {pretrained_model_name_or_path} " + "but there is a file for PyTorch weights." + ) + else: + raise EnvironmentError( + f"Error no file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME} found in directory " + f"{pretrained_model_name_or_path}." + ) + else: + try: + model_file = hf_hub_download( + pretrained_model_name_or_path, + filename=FLAX_WEIGHTS_NAME, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + resume_download=resume_download, + local_files_only=local_files_only, + use_auth_token=use_auth_token, + user_agent=user_agent, + subfolder=subfolder, + revision=revision, + ) + + except RepositoryNotFoundError: + raise EnvironmentError( + f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier " + "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a " + "token having permission to this repo with `use_auth_token` or log in with `huggingface-cli " + "login` and pass `use_auth_token=True`." + ) + except RevisionNotFoundError: + raise EnvironmentError( + f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for " + "this model name. Check the model page at " + f"'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions." + ) + except EntryNotFoundError: + raise EnvironmentError( + f"{pretrained_model_name_or_path} does not appear to have a file named {FLAX_WEIGHTS_NAME}." + ) + except HTTPError as err: + raise EnvironmentError( + f"There was a specific connection error when trying to load {pretrained_model_name_or_path}:\n" + f"{err}" + ) + except ValueError: + raise EnvironmentError( + f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it" + f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a" + f" directory containing a file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME}.\nCheckout your" + " internet connection or see how to run the library in offline mode at" + " 'https://huggingface.co/docs/transformers/installation#offline-mode'." + ) + except EnvironmentError: + raise EnvironmentError( + f"Can't load the model for '{pretrained_model_name_or_path}'. If you were trying to load it from " + "'https://huggingface.co/models', make sure you don't have a local directory with the same name. " + f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory " + f"containing a file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME}." + ) + + try: + with open(model_file, "rb") as state_f: + state = from_bytes(cls, state_f.read()) + except (UnpicklingError, msgpack.exceptions.ExtraData) as e: + try: + with open(model_file) as f: + if f.read().startswith("version"): + raise OSError( + "You seem to have cloned a repository without having git-lfs installed. Please" + " install git-lfs and run `git lfs install` followed by `git lfs pull` in the" + " folder you cloned." + ) + else: + raise ValueError from e + except (UnicodeDecodeError, ValueError): + raise EnvironmentError(f"Unable to convert {model_file} to Flax deserializable object. ") + # make sure all arrays are stored as jnp.arrays + # NOTE: This is to prevent a bug this will be fixed in Flax >= v0.3.4: + # https://github.com/google/flax/issues/1261 + state = jax.tree_util.tree_map(lambda x: jax.device_put(x, jax.devices("cpu")[0]), state) + + # flatten dicts + state = flatten_dict(state) + + # dictionary of key: dtypes for the model params + param_dtypes = jax.tree_map(lambda x: x.dtype, state) + # extract keys of parameters not in jnp.float32 + fp16_params = [k for k in param_dtypes if param_dtypes[k] == jnp.float16] + bf16_params = [k for k in param_dtypes if param_dtypes[k] == jnp.bfloat16] + + # raise a warning if any of the parameters are not in jnp.float32 + if len(fp16_params) > 0: + logger.warning( + f"Some of the weights of {model.__class__.__name__} were initialized in float16 precision from " + f"the model checkpoint at {pretrained_model_name_or_path}:\n{fp16_params}\n" + "You should probably UPCAST the model weights to float32 if this was not intended. " + "See [`~ModelMixin.to_fp32`] for further information on how to do this." + ) + + if len(bf16_params) > 0: + logger.warning( + f"Some of the weights of {model.__class__.__name__} were initialized in bfloat16 precision from " + f"the model checkpoint at {pretrained_model_name_or_path}:\n{bf16_params}\n" + "You should probably UPCAST the model weights to float32 if this was not intended. " + "See [`~ModelMixin.to_fp32`] for further information on how to do this." + ) + + return model, unflatten_dict(state) + + def save_pretrained( + self, + save_directory: Union[str, os.PathLike], + params: Union[Dict, FrozenDict], + is_main_process: bool = True, + ): + """ + Save a model and its configuration file to a directory, so that it can be re-loaded using the + `[`~FlaxModelMixin.from_pretrained`]` class method + + Arguments: + save_directory (`str` or `os.PathLike`): + Directory to which to save. 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 when in distributed training like + TPUs and 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. + """ + if os.path.isfile(save_directory): + logger.error(f"Provided path ({save_directory}) should be a directory, not a file") + return + + os.makedirs(save_directory, exist_ok=True) + + model_to_save = self + + # Attach architecture to the config + # Save the config + if is_main_process: + model_to_save.save_config(save_directory) + + # save model + output_model_file = os.path.join(save_directory, FLAX_WEIGHTS_NAME) + with open(output_model_file, "wb") as f: + model_bytes = to_bytes(params) + f.write(model_bytes) + + logger.info(f"Model weights saved in {output_model_file}") diff --git a/src/diffusers/utils/dummy_flax_objects.py b/src/diffusers/utils/dummy_flax_objects.py index b5f4362bcb6e..981dc5586ad9 100644 --- a/src/diffusers/utils/dummy_flax_objects.py +++ b/src/diffusers/utils/dummy_flax_objects.py @@ -4,6 +4,13 @@ from ..utils import DummyObject, requires_backends +class FlaxModelMixin(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + class FlaxPNDMScheduler(metaclass=DummyObject): _backends = ["flax"]