up

src/diffusers/models/modeling_utils.py

@@ -675,6 +675,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         variant: Optional[str] = None,
         max_shard_size: Union[int, str] = "10GB",
         push_to_hub: bool = False,
+        use_flashpack: bool = False,
         **kwargs,
     ):
         """
@@ -707,6 +708,9 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 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).
+            use_flashpack (`bool`, *optional*, defaults to `False`):
+                Whether to save the model in [FlashPack](https://github.com/fal-ai/flashpack) format. FlashPack is a
+                binary format that allows for faster loading. Requires the `flashpack` library to be installed.
             kwargs (`Dict[str, Any]`, *optional*):
                 Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
         """
@@ -727,12 +731,6 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                     " 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:
@@ -746,67 +744,80 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         # 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()
+        if use_flashpack:
+            if not is_main_process:
+                return
 
-        # 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
-        )
+            from ..utils.flashpack_utils import save_flashpack
 
-        # 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}."
-            )
+            save_flashpack(model_to_save, save_directory, variant=variant)
         else:
-            path_to_weights = os.path.join(save_directory, weights_name)
-            logger.info(f"Model weights saved in {path_to_weights}")
+            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"
+            )
 
+            state_dict = model_to_save.state_dict()
+            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)
+
+            # Save each shard
+            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)
+
+            # Save index file if sharded
+            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}")
+
+        # Push to hub if requested (common to both paths)
         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)
@@ -939,6 +950,10 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 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_flashpack (`bool`, *optional*, defaults to `False`):
+                If set to `True`, the model is first loaded from `flashpack` (https://github.com/fal-ai/flashpack)
+                weights if a compatible `.flashpack` file is found. If flashpack is unavailable or the `.flashpack`
+                file cannot be used, automatic fallback to the standard loading path (for example, `safetensors`).
             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.
@@ -982,6 +997,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         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_flashpack = kwargs.pop("use_flashpack", False)
         quantization_config = kwargs.pop("quantization_config", None)
         dduf_entries: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_entries", None)
         disable_mmap = kwargs.pop("disable_mmap", False)
@@ -1199,7 +1215,31 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             from .modeling_pytorch_flax_utils import load_flax_checkpoint_in_pytorch_model
 
             model = load_flax_checkpoint_in_pytorch_model(model, resolved_model_file)
-        else:
+
+        flashpack_file = None
+        if use_flashpack:
+            try:
+                flashpack_file = _get_model_file(
+                    pretrained_model_name_or_path,
+                    weights_name=_add_variant("model.flashpack", variant),
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                    commit_hash=commit_hash,
+                    dduf_entries=dduf_entries,
+                )
+            except EnvironmentError:
+                flashpack_file = None
+                logger.warning(
+                    "`use_flashpack` was specified to be True but not flashpack file was found. Resorting to non-flashpack alternatives."
+                )
+
+        if flashpack_file is None:
             # in the case it is sharded, we have already the index
             if is_sharded:
                 resolved_model_file, sharded_metadata = _get_checkpoint_shard_files(
@@ -1215,6 +1255,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                     dduf_entries=dduf_entries,
                 )
             elif use_safetensors:
+                logger.warning("Trying to load model weights with safetensors format.")
                 try:
                     resolved_model_file = _get_model_file(
                         pretrained_model_name_or_path,
@@ -1280,6 +1321,29 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         if dtype_orig is not None:
             torch.set_default_dtype(dtype_orig)
 
+        if flashpack_file is not None:
+            from ..utils.flashpack_utils import load_flashpack
+
+            # Even when using FlashPack, we preserve `low_cpu_mem_usage` behavior by initializing
+            # the model with meta tensors. Since FlashPack cannot write into meta tensors, we
+            # explicitly materialize parameters before loading to ensure correctness and parity
+            # with the standard loading path.
+            if any(p.device.type == "meta" for p in model.parameters()):
+                model.to_empty(device="cpu")
+            load_flashpack(model, flashpack_file)
+            model.register_to_config(_name_or_path=pretrained_model_name_or_path)
+            model.eval()
+
+            if output_loading_info:
+                return model, {
+                    "missing_keys": [],
+                    "unexpected_keys": [],
+                    "mismatched_keys": [],
+                    "error_msgs": [],
+                }
+
+            return model
+
         state_dict = None
         if not is_sharded:
             # Time to load the checkpoint
@@ -1327,7 +1391,6 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             keep_in_fp32_modules=keep_in_fp32_modules,
             dduf_entries=dduf_entries,
             is_parallel_loading_enabled=is_parallel_loading_enabled,
-            disable_mmap=disable_mmap,
         )
         loading_info = {
             "missing_keys": missing_keys,
@@ -1373,6 +1436,8 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         if output_loading_info:
             return model, loading_info
 
+        logger.warning(f"Model till end {pretrained_model_name_or_path} loaded successfully")
+
         return model
 
     # Adapted from `transformers`.

src/diffusers/pipelines/pipeline_loading_utils.py

@@ -756,6 +756,7 @@ def load_sub_model(
     low_cpu_mem_usage: bool,
     cached_folder: Union[str, os.PathLike],
     use_safetensors: bool,
+    use_flashpack: bool,
     dduf_entries: Optional[Dict[str, DDUFEntry]],
     provider_options: Any,
     disable_mmap: bool,
@@ -838,6 +839,9 @@ def load_sub_model(
         loading_kwargs["variant"] = model_variants.pop(name, None)
         loading_kwargs["use_safetensors"] = use_safetensors
 
+        if is_diffusers_model:
+            loading_kwargs["use_flashpack"] = use_flashpack
+
         if from_flax:
             loading_kwargs["from_flax"] = True
 

src/diffusers/pipelines/pipeline_utils.py

@@ -243,6 +243,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         variant: Optional[str] = None,
         max_shard_size: Optional[Union[int, str]] = None,
         push_to_hub: bool = False,
+        use_flashpack: bool = False,
         **kwargs,
     ):
         """
@@ -268,7 +269,9 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 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).
-
+            use_flashpack (`bool`, *optional*, defaults to `False`):
+                Whether or not to use `flashpack` to save the model weights. Requires the `flashpack` library: `pip
+                install flashpack`.
             kwargs (`Dict[str, Any]`, *optional*):
                 Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
         """
@@ -340,6 +343,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             save_method_accept_safe = "safe_serialization" in save_method_signature.parameters
             save_method_accept_variant = "variant" in save_method_signature.parameters
             save_method_accept_max_shard_size = "max_shard_size" in save_method_signature.parameters
+            save_method_accept_flashpack = "use_flashpack" in save_method_signature.parameters
 
             save_kwargs = {}
             if save_method_accept_safe:
@@ -349,6 +353,8 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             if save_method_accept_max_shard_size and max_shard_size is not None:
                 # max_shard_size is expected to not be None in ModelMixin
                 save_kwargs["max_shard_size"] = max_shard_size
+            if save_method_accept_flashpack:
+                save_kwargs["use_flashpack"] = use_flashpack
 
             save_method(os.path.join(save_directory, pipeline_component_name), **save_kwargs)
 
@@ -707,6 +713,11 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 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_flashpack (`bool`, *optional*, defaults to `False`):
+                If set to `True`, the model is first loaded from `flashpack` weights if a compatible `.flashpack` file
+                is found. If flashpack is unavailable or the `.flashpack` file cannot be used, automatic fallback to
+                the standard loading path (for example, `safetensors`). Requires the `flashpack` library: `pip install
+                flashpack`.
             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
@@ -772,6 +783,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         variant = kwargs.pop("variant", None)
         dduf_file = kwargs.pop("dduf_file", None)
         use_safetensors = kwargs.pop("use_safetensors", None)
+        use_flashpack = kwargs.pop("use_flashpack", False)
         use_onnx = kwargs.pop("use_onnx", None)
         load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)
         quantization_config = kwargs.pop("quantization_config", None)
@@ -1061,6 +1073,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                     low_cpu_mem_usage=low_cpu_mem_usage,
                     cached_folder=cached_folder,
                     use_safetensors=use_safetensors,
+                    use_flashpack=use_flashpack,
                     dduf_entries=dduf_entries,
                     provider_options=provider_options,
                     disable_mmap=disable_mmap,

src/diffusers/utils/flashpack_utils.py

@@ -0,0 +1,81 @@
+import json
+import os
+from typing import Optional
+
+from ..utils import _add_variant
+from .import_utils import is_flashpack_available
+from .logging import get_logger
+
+
+logger = get_logger(__name__)
+
+
+def save_flashpack(
+    model,
+    save_directory: str,
+    variant: Optional[str] = None,
+    is_main_process: bool = True,
+):
+    """
+    Save model weights in FlashPack format along with a metadata config.
+
+    Args:
+        model: Diffusers model instance
+        save_directory (`str`): Directory to save weights
+        variant (`str`, *optional*): Model variant
+    """
+    if not is_flashpack_available():
+        raise ImportError(
+            "The `use_flashpack=True` argument requires the `flashpack` package. "
+            "Install it with `pip install flashpack`."
+        )
+
+    from flashpack import pack_to_file
+
+    os.makedirs(save_directory, exist_ok=True)
+
+    weights_name = _add_variant("model.flashpack", variant)
+    weights_path = os.path.join(save_directory, weights_name)
+    config_path = os.path.join(save_directory, "flashpack_config.json")
+
+    try:
+        target_dtype = getattr(model, "dtype", None)
+        logger.warning(f"Dtype used for FlashPack save: {target_dtype}")
+
+        # 1. Save binary weights
+        pack_to_file(model, weights_path, target_dtype=target_dtype)
+
+        # 2. Save config metadata (best-effort)
+        if hasattr(model, "config"):
+            try:
+                if hasattr(model.config, "to_dict"):
+                    config_data = model.config.to_dict()
+                else:
+                    config_data = dict(model.config)
+
+                with open(config_path, "w") as f:
+                    json.dump(config_data, f, indent=4)
+
+            except Exception as config_err:
+                logger.warning(f"FlashPack weights saved, but config serialization failed: {config_err}")
+
+    except Exception as e:
+        logger.error(f"Failed to save weights in FlashPack format: {e}")
+        raise
+
+
+def load_flashpack(model, flashpack_file: str):
+    """
+    Assign FlashPack weights from a file into an initialized PyTorch model.
+    """
+    if not is_flashpack_available():
+        raise ImportError("FlashPack weights require the `flashpack` package. Install with `pip install flashpack`.")
+
+    from flashpack import assign_from_file
+
+    logger.warning(f"Loading FlashPack weights from {flashpack_file}")
+
+    try:
+        assign_from_file(model, flashpack_file)
+    except Exception as e:
+        raise RuntimeError(f"Failed to load FlashPack weights from {flashpack_file}") from e

src/diffusers/utils/import_utils.py

@@ -231,6 +231,7 @@ _aiter_available, _aiter_version = _is_package_available("aiter")
 _kornia_available, _kornia_version = _is_package_available("kornia")
 _nvidia_modelopt_available, _nvidia_modelopt_version = _is_package_available("modelopt", get_dist_name=True)
 _av_available, _av_version = _is_package_available("av")
+_flashpack_available, _flashpack_version = _is_package_available("flashpack")
 
 
 def is_torch_available():
@@ -425,6 +426,10 @@ def is_av_available():
     return _av_available
 
 
+def is_flashpack_available():
+    return _flashpack_available
+
+
 # docstyle-ignore
 FLAX_IMPORT_ERROR = """
 {0} requires the FLAX library but it was not found in your environment. Checkout the instructions on the
@@ -942,6 +947,16 @@ def is_aiter_version(operation: str, version: str):
     return compare_versions(parse(_aiter_version), operation, version)
 
 
+@cache
+def is_flashpack_version(operation: str, version: str):
+    """
+    Compares the current flashpack version to a given reference with an operation.
+    """
+    if not _flashpack_available:
+        return False
+    return compare_versions(parse(_flashpack_version), operation, version)
+
+
 def get_objects_from_module(module):
     """
     Returns a dict of object names and values in a module, while skipping private/internal objects