Release: v0.32.2

* update

* update

* update

* update

---------

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

.github/workflows/push_tests.yml

@@ -83,7 +83,7 @@ jobs:
           python utils/print_env.py
       - name: PyTorch CUDA checkpoint tests on Ubuntu
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
           # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
@@ -137,7 +137,7 @@ jobs:
 
     - name: Run PyTorch CUDA tests
       env:
-        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
         # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
         CUBLAS_WORKSPACE_CONFIG: :16:8
       run: |

docs/source/en/api/pipelines/flux.md

@@ -305,6 +305,10 @@ image = control_pipe(
 image.save("output.png")
 ```
 
+## Note about `unload_lora_weights()` when using Flux LoRAs
+
+When unloading the Control LoRA weights, call `pipe.unload_lora_weights(reset_to_overwritten_params=True)` to reset the `pipe.transformer` completely back to its original form. The resultant pipeline can then be used with methods like [`DiffusionPipeline.from_pipe`]. More details about this argument are available in [this PR](https://github.com/huggingface/diffusers/pull/10397).
+
 ## Running FP16 inference
 
 Flux can generate high-quality images with FP16 (i.e. to accelerate inference on Turing/Volta GPUs) but produces different outputs compared to FP32/BF16. The issue is that some activations in the text encoders have to be clipped when running in FP16, which affects the overall image. Forcing text encoders to run with FP32 inference thus removes this output difference. See [here](https://github.com/huggingface/diffusers/pull/9097#issuecomment-2272292516) for details.

setup.py

@@ -254,7 +254,7 @@ version_range_max = max(sys.version_info[1], 10) + 1
 
 setup(
     name="diffusers",
-    version="0.32.1",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="0.32.2",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     description="State-of-the-art diffusion in PyTorch and JAX.",
     long_description=open("README.md", "r", encoding="utf-8").read(),
     long_description_content_type="text/markdown",

src/diffusers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.32.1"
+__version__ = "0.32.2"
 
 from typing import TYPE_CHECKING
 

src/diffusers/loaders/lora_base.py

@@ -28,13 +28,20 @@ from ..models.modeling_utils import ModelMixin, load_state_dict
 from ..utils import (
     USE_PEFT_BACKEND,
     _get_model_file,
+    convert_state_dict_to_diffusers,
+    convert_state_dict_to_peft,
     delete_adapter_layers,
     deprecate,
+    get_adapter_name,
+    get_peft_kwargs,
     is_accelerate_available,
     is_peft_available,
+    is_peft_version,
     is_transformers_available,
+    is_transformers_version,
     logging,
     recurse_remove_peft_layers,
+    scale_lora_layers,
     set_adapter_layers,
     set_weights_and_activate_adapters,
 )
@@ -43,6 +50,8 @@ from ..utils import (
 if is_transformers_available():
     from transformers import PreTrainedModel
 
+    from ..models.lora import text_encoder_attn_modules, text_encoder_mlp_modules
+
 if is_peft_available():
     from peft.tuners.tuners_utils import BaseTunerLayer
 
@@ -297,6 +306,152 @@ def _best_guess_weight_name(
     return weight_name
 
 
+def _load_lora_into_text_encoder(
+    state_dict,
+    network_alphas,
+    text_encoder,
+    prefix=None,
+    lora_scale=1.0,
+    text_encoder_name="text_encoder",
+    adapter_name=None,
+    _pipeline=None,
+    low_cpu_mem_usage=False,
+):
+    if not USE_PEFT_BACKEND:
+        raise ValueError("PEFT backend is required for this method.")
+
+    peft_kwargs = {}
+    if low_cpu_mem_usage:
+        if not is_peft_version(">=", "0.13.1"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+        if not is_transformers_version(">", "4.45.2"):
+            # Note from sayakpaul: It's not in `transformers` stable yet.
+            # https://github.com/huggingface/transformers/pull/33725/
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `transformers` version. Please update it with `pip install -U transformers`."
+            )
+        peft_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
+
+    from peft import LoraConfig
+
+    # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
+    # then the `state_dict` keys should have `unet_name` and/or `text_encoder_name` as
+    # their prefixes.
+    keys = list(state_dict.keys())
+    prefix = text_encoder_name if prefix is None else prefix
+
+    # Safe prefix to check with.
+    if any(text_encoder_name in key for key in keys):
+        # Load the layers corresponding to text encoder and make necessary adjustments.
+        text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
+        text_encoder_lora_state_dict = {
+            k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
+        }
+
+        if len(text_encoder_lora_state_dict) > 0:
+            logger.info(f"Loading {prefix}.")
+            rank = {}
+            text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)
+
+            # convert state dict
+            text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)
+
+            for name, _ in text_encoder_attn_modules(text_encoder):
+                for module in ("out_proj", "q_proj", "k_proj", "v_proj"):
+                    rank_key = f"{name}.{module}.lora_B.weight"
+                    if rank_key not in text_encoder_lora_state_dict:
+                        continue
+                    rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]
+
+            for name, _ in text_encoder_mlp_modules(text_encoder):
+                for module in ("fc1", "fc2"):
+                    rank_key = f"{name}.{module}.lora_B.weight"
+                    if rank_key not in text_encoder_lora_state_dict:
+                        continue
+                    rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]
+
+            if network_alphas is not None:
+                alpha_keys = [k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix]
+                network_alphas = {k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys}
+
+            lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
+
+            if "use_dora" in lora_config_kwargs:
+                if lora_config_kwargs["use_dora"]:
+                    if is_peft_version("<", "0.9.0"):
+                        raise ValueError(
+                            "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
+                        )
+                else:
+                    if is_peft_version("<", "0.9.0"):
+                        lora_config_kwargs.pop("use_dora")
+
+            if "lora_bias" in lora_config_kwargs:
+                if lora_config_kwargs["lora_bias"]:
+                    if is_peft_version("<=", "0.13.2"):
+                        raise ValueError(
+                            "You need `peft` 0.14.0 at least to use `bias` in LoRAs. Please upgrade your installation of `peft`."
+                        )
+                else:
+                    if is_peft_version("<=", "0.13.2"):
+                        lora_config_kwargs.pop("lora_bias")
+
+            lora_config = LoraConfig(**lora_config_kwargs)
+
+            # adapter_name
+            if adapter_name is None:
+                adapter_name = get_adapter_name(text_encoder)
+
+            is_model_cpu_offload, is_sequential_cpu_offload = _func_optionally_disable_offloading(_pipeline)
+
+            # inject LoRA layers and load the state dict
+            # in transformers we automatically check whether the adapter name is already in use or not
+            text_encoder.load_adapter(
+                adapter_name=adapter_name,
+                adapter_state_dict=text_encoder_lora_state_dict,
+                peft_config=lora_config,
+                **peft_kwargs,
+            )
+
+            # scale LoRA layers with `lora_scale`
+            scale_lora_layers(text_encoder, weight=lora_scale)
+
+            text_encoder.to(device=text_encoder.device, dtype=text_encoder.dtype)
+
+            # Offload back.
+            if is_model_cpu_offload:
+                _pipeline.enable_model_cpu_offload()
+            elif is_sequential_cpu_offload:
+                _pipeline.enable_sequential_cpu_offload()
+            # Unsafe code />
+
+
+def _func_optionally_disable_offloading(_pipeline):
+    is_model_cpu_offload = False
+    is_sequential_cpu_offload = False
+
+    if _pipeline is not None and _pipeline.hf_device_map is None:
+        for _, component in _pipeline.components.items():
+            if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
+                if not is_model_cpu_offload:
+                    is_model_cpu_offload = isinstance(component._hf_hook, CpuOffload)
+                if not is_sequential_cpu_offload:
+                    is_sequential_cpu_offload = (
+                        isinstance(component._hf_hook, AlignDevicesHook)
+                        or hasattr(component._hf_hook, "hooks")
+                        and isinstance(component._hf_hook.hooks[0], AlignDevicesHook)
+                    )
+
+                logger.info(
+                    "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+                )
+                remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+    return (is_model_cpu_offload, is_sequential_cpu_offload)
+
+
 class LoraBaseMixin:
     """Utility class for handling LoRAs."""
 
@@ -327,27 +482,7 @@ class LoraBaseMixin:
             tuple:
                 A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
         """
-        is_model_cpu_offload = False
-        is_sequential_cpu_offload = False
-
-        if _pipeline is not None and _pipeline.hf_device_map is None:
-            for _, component in _pipeline.components.items():
-                if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
-                    if not is_model_cpu_offload:
-                        is_model_cpu_offload = isinstance(component._hf_hook, CpuOffload)
-                    if not is_sequential_cpu_offload:
-                        is_sequential_cpu_offload = (
-                            isinstance(component._hf_hook, AlignDevicesHook)
-                            or hasattr(component._hf_hook, "hooks")
-                            and isinstance(component._hf_hook.hooks[0], AlignDevicesHook)
-                        )
-
-                    logger.info(
-                        "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
-                    )
-                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
-
-        return (is_model_cpu_offload, is_sequential_cpu_offload)
+        return _func_optionally_disable_offloading(_pipeline=_pipeline)
 
     @classmethod
     def _fetch_state_dict(cls, *args, **kwargs):

src/diffusers/loaders/lora_conversion_utils.py

@@ -973,3 +973,178 @@ def _convert_bfl_flux_control_lora_to_diffusers(original_state_dict):
         converted_state_dict[f"transformer.{key}"] = converted_state_dict.pop(key)
 
     return converted_state_dict
+
+
+def _convert_hunyuan_video_lora_to_diffusers(original_state_dict):
+    converted_state_dict = {k: original_state_dict.pop(k) for k in list(original_state_dict.keys())}
+
+    def remap_norm_scale_shift_(key, state_dict):
+        weight = state_dict.pop(key)
+        shift, scale = weight.chunk(2, dim=0)
+        new_weight = torch.cat([scale, shift], dim=0)
+        state_dict[key.replace("final_layer.adaLN_modulation.1", "norm_out.linear")] = new_weight
+
+    def remap_txt_in_(key, state_dict):
+        def rename_key(key):
+            new_key = key.replace("individual_token_refiner.blocks", "token_refiner.refiner_blocks")
+            new_key = new_key.replace("adaLN_modulation.1", "norm_out.linear")
+            new_key = new_key.replace("txt_in", "context_embedder")
+            new_key = new_key.replace("t_embedder.mlp.0", "time_text_embed.timestep_embedder.linear_1")
+            new_key = new_key.replace("t_embedder.mlp.2", "time_text_embed.timestep_embedder.linear_2")
+            new_key = new_key.replace("c_embedder", "time_text_embed.text_embedder")
+            new_key = new_key.replace("mlp", "ff")
+            return new_key
+
+        if "self_attn_qkv" in key:
+            weight = state_dict.pop(key)
+            to_q, to_k, to_v = weight.chunk(3, dim=0)
+            state_dict[rename_key(key.replace("self_attn_qkv", "attn.to_q"))] = to_q
+            state_dict[rename_key(key.replace("self_attn_qkv", "attn.to_k"))] = to_k
+            state_dict[rename_key(key.replace("self_attn_qkv", "attn.to_v"))] = to_v
+        else:
+            state_dict[rename_key(key)] = state_dict.pop(key)
+
+    def remap_img_attn_qkv_(key, state_dict):
+        weight = state_dict.pop(key)
+        if "lora_A" in key:
+            state_dict[key.replace("img_attn_qkv", "attn.to_q")] = weight
+            state_dict[key.replace("img_attn_qkv", "attn.to_k")] = weight
+            state_dict[key.replace("img_attn_qkv", "attn.to_v")] = weight
+        else:
+            to_q, to_k, to_v = weight.chunk(3, dim=0)
+            state_dict[key.replace("img_attn_qkv", "attn.to_q")] = to_q
+            state_dict[key.replace("img_attn_qkv", "attn.to_k")] = to_k
+            state_dict[key.replace("img_attn_qkv", "attn.to_v")] = to_v
+
+    def remap_txt_attn_qkv_(key, state_dict):
+        weight = state_dict.pop(key)
+        if "lora_A" in key:
+            state_dict[key.replace("txt_attn_qkv", "attn.add_q_proj")] = weight
+            state_dict[key.replace("txt_attn_qkv", "attn.add_k_proj")] = weight
+            state_dict[key.replace("txt_attn_qkv", "attn.add_v_proj")] = weight
+        else:
+            to_q, to_k, to_v = weight.chunk(3, dim=0)
+            state_dict[key.replace("txt_attn_qkv", "attn.add_q_proj")] = to_q
+            state_dict[key.replace("txt_attn_qkv", "attn.add_k_proj")] = to_k
+            state_dict[key.replace("txt_attn_qkv", "attn.add_v_proj")] = to_v
+
+    def remap_single_transformer_blocks_(key, state_dict):
+        hidden_size = 3072
+
+        if "linear1.lora_A.weight" in key or "linear1.lora_B.weight" in key:
+            linear1_weight = state_dict.pop(key)
+            if "lora_A" in key:
+                new_key = key.replace("single_blocks", "single_transformer_blocks").removesuffix(
+                    ".linear1.lora_A.weight"
+                )
+                state_dict[f"{new_key}.attn.to_q.lora_A.weight"] = linear1_weight
+                state_dict[f"{new_key}.attn.to_k.lora_A.weight"] = linear1_weight
+                state_dict[f"{new_key}.attn.to_v.lora_A.weight"] = linear1_weight
+                state_dict[f"{new_key}.proj_mlp.lora_A.weight"] = linear1_weight
+            else:
+                split_size = (hidden_size, hidden_size, hidden_size, linear1_weight.size(0) - 3 * hidden_size)
+                q, k, v, mlp = torch.split(linear1_weight, split_size, dim=0)
+                new_key = key.replace("single_blocks", "single_transformer_blocks").removesuffix(
+                    ".linear1.lora_B.weight"
+                )
+                state_dict[f"{new_key}.attn.to_q.lora_B.weight"] = q
+                state_dict[f"{new_key}.attn.to_k.lora_B.weight"] = k
+                state_dict[f"{new_key}.attn.to_v.lora_B.weight"] = v
+                state_dict[f"{new_key}.proj_mlp.lora_B.weight"] = mlp
+
+        elif "linear1.lora_A.bias" in key or "linear1.lora_B.bias" in key:
+            linear1_bias = state_dict.pop(key)
+            if "lora_A" in key:
+                new_key = key.replace("single_blocks", "single_transformer_blocks").removesuffix(
+                    ".linear1.lora_A.bias"
+                )
+                state_dict[f"{new_key}.attn.to_q.lora_A.bias"] = linear1_bias
+                state_dict[f"{new_key}.attn.to_k.lora_A.bias"] = linear1_bias
+                state_dict[f"{new_key}.attn.to_v.lora_A.bias"] = linear1_bias
+                state_dict[f"{new_key}.proj_mlp.lora_A.bias"] = linear1_bias
+            else:
+                split_size = (hidden_size, hidden_size, hidden_size, linear1_bias.size(0) - 3 * hidden_size)
+                q_bias, k_bias, v_bias, mlp_bias = torch.split(linear1_bias, split_size, dim=0)
+                new_key = key.replace("single_blocks", "single_transformer_blocks").removesuffix(
+                    ".linear1.lora_B.bias"
+                )
+                state_dict[f"{new_key}.attn.to_q.lora_B.bias"] = q_bias
+                state_dict[f"{new_key}.attn.to_k.lora_B.bias"] = k_bias
+                state_dict[f"{new_key}.attn.to_v.lora_B.bias"] = v_bias
+                state_dict[f"{new_key}.proj_mlp.lora_B.bias"] = mlp_bias
+
+        else:
+            new_key = key.replace("single_blocks", "single_transformer_blocks")
+            new_key = new_key.replace("linear2", "proj_out")
+            new_key = new_key.replace("q_norm", "attn.norm_q")
+            new_key = new_key.replace("k_norm", "attn.norm_k")
+            state_dict[new_key] = state_dict.pop(key)
+
+    TRANSFORMER_KEYS_RENAME_DICT = {
+        "img_in": "x_embedder",
+        "time_in.mlp.0": "time_text_embed.timestep_embedder.linear_1",
+        "time_in.mlp.2": "time_text_embed.timestep_embedder.linear_2",
+        "guidance_in.mlp.0": "time_text_embed.guidance_embedder.linear_1",
+        "guidance_in.mlp.2": "time_text_embed.guidance_embedder.linear_2",
+        "vector_in.in_layer": "time_text_embed.text_embedder.linear_1",
+        "vector_in.out_layer": "time_text_embed.text_embedder.linear_2",
+        "double_blocks": "transformer_blocks",
+        "img_attn_q_norm": "attn.norm_q",
+        "img_attn_k_norm": "attn.norm_k",
+        "img_attn_proj": "attn.to_out.0",
+        "txt_attn_q_norm": "attn.norm_added_q",
+        "txt_attn_k_norm": "attn.norm_added_k",
+        "txt_attn_proj": "attn.to_add_out",
+        "img_mod.linear": "norm1.linear",
+        "img_norm1": "norm1.norm",
+        "img_norm2": "norm2",
+        "img_mlp": "ff",
+        "txt_mod.linear": "norm1_context.linear",
+        "txt_norm1": "norm1.norm",
+        "txt_norm2": "norm2_context",
+        "txt_mlp": "ff_context",
+        "self_attn_proj": "attn.to_out.0",
+        "modulation.linear": "norm.linear",
+        "pre_norm": "norm.norm",
+        "final_layer.norm_final": "norm_out.norm",
+        "final_layer.linear": "proj_out",
+        "fc1": "net.0.proj",
+        "fc2": "net.2",
+        "input_embedder": "proj_in",
+    }
+
+    TRANSFORMER_SPECIAL_KEYS_REMAP = {
+        "txt_in": remap_txt_in_,
+        "img_attn_qkv": remap_img_attn_qkv_,
+        "txt_attn_qkv": remap_txt_attn_qkv_,
+        "single_blocks": remap_single_transformer_blocks_,
+        "final_layer.adaLN_modulation.1": remap_norm_scale_shift_,
+    }
+
+    # Some folks attempt to make their state dict compatible with diffusers by adding "transformer." prefix to all keys
+    # and use their custom code. To make sure both "original" and "attempted diffusers" loras work as expected, we make
+    # sure that both follow the same initial format by stripping off the "transformer." prefix.
+    for key in list(converted_state_dict.keys()):
+        if key.startswith("transformer."):
+            converted_state_dict[key[len("transformer.") :]] = converted_state_dict.pop(key)
+        if key.startswith("diffusion_model."):
+            converted_state_dict[key[len("diffusion_model.") :]] = converted_state_dict.pop(key)
+
+    # Rename and remap the state dict keys
+    for key in list(converted_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        converted_state_dict[new_key] = converted_state_dict.pop(key)
+
+    for key in list(converted_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, converted_state_dict)
+
+    # Add back the "transformer." prefix
+    for key in list(converted_state_dict.keys()):
+        converted_state_dict[f"transformer.{key}"] = converted_state_dict.pop(key)
+
+    return converted_state_dict

src/diffusers/loaders/peft.py

@@ -20,7 +20,6 @@ from typing import Dict, List, Optional, Union
 
 import safetensors
 import torch
-import torch.nn as nn
 
 from ..utils import (
     MIN_PEFT_VERSION,
@@ -30,20 +29,16 @@ from ..utils import (
     delete_adapter_layers,
     get_adapter_name,
     get_peft_kwargs,
-    is_accelerate_available,
     is_peft_available,
     is_peft_version,
     logging,
     set_adapter_layers,
     set_weights_and_activate_adapters,
 )
-from .lora_base import _fetch_state_dict
+from .lora_base import _fetch_state_dict, _func_optionally_disable_offloading
 from .unet_loader_utils import _maybe_expand_lora_scales
 
 
-if is_accelerate_available():
-    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
-
 logger = logging.get_logger(__name__)
 
 _SET_ADAPTER_SCALE_FN_MAPPING = {
@@ -140,27 +135,7 @@ class PeftAdapterMixin:
             tuple:
                 A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
         """
-        is_model_cpu_offload = False
-        is_sequential_cpu_offload = False
-
-        if _pipeline is not None and _pipeline.hf_device_map is None:
-            for _, component in _pipeline.components.items():
-                if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
-                    if not is_model_cpu_offload:
-                        is_model_cpu_offload = isinstance(component._hf_hook, CpuOffload)
-                    if not is_sequential_cpu_offload:
-                        is_sequential_cpu_offload = (
-                            isinstance(component._hf_hook, AlignDevicesHook)
-                            or hasattr(component._hf_hook, "hooks")
-                            and isinstance(component._hf_hook.hooks[0], AlignDevicesHook)
-                        )
-
-                    logger.info(
-                        "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
-                    )
-                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
-
-        return (is_model_cpu_offload, is_sequential_cpu_offload)
+        return _func_optionally_disable_offloading(_pipeline=_pipeline)
 
     def load_lora_adapter(self, pretrained_model_name_or_path_or_dict, prefix="transformer", **kwargs):
         r"""

src/diffusers/loaders/single_file_utils.py

@@ -595,10 +595,14 @@ def infer_diffusers_model_type(checkpoint):
         if any(
             g in checkpoint for g in ["guidance_in.in_layer.bias", "model.diffusion_model.guidance_in.in_layer.bias"]
         ):
-            if checkpoint["img_in.weight"].shape[1] == 384:
-                model_type = "flux-fill"
+            if "model.diffusion_model.img_in.weight" in checkpoint:
+                key = "model.diffusion_model.img_in.weight"
+            else:
+                key = "img_in.weight"
 
-            elif checkpoint["img_in.weight"].shape[1] == 128:
+            if checkpoint[key].shape[1] == 384:
+                model_type = "flux-fill"
+            elif checkpoint[key].shape[1] == 128:
                 model_type = "flux-depth"
             else:
                 model_type = "flux-dev"

src/diffusers/loaders/unet.py

@@ -21,7 +21,6 @@ import safetensors
 import torch
 import torch.nn.functional as F
 from huggingface_hub.utils import validate_hf_hub_args
-from torch import nn
 
 from ..models.embeddings import (
     ImageProjection,
@@ -44,13 +43,11 @@ from ..utils import (
     is_torch_version,
     logging,
 )
+from .lora_base import _func_optionally_disable_offloading
 from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE, TEXT_ENCODER_NAME, UNET_NAME
 from .utils import AttnProcsLayers
 
 
-if is_accelerate_available():
-    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
-
 logger = logging.get_logger(__name__)
 
 
@@ -400,27 +397,7 @@ class UNet2DConditionLoadersMixin:
             tuple:
                 A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
         """
-        is_model_cpu_offload = False
-        is_sequential_cpu_offload = False
-
-        if _pipeline is not None and _pipeline.hf_device_map is None:
-            for _, component in _pipeline.components.items():
-                if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
-                    if not is_model_cpu_offload:
-                        is_model_cpu_offload = isinstance(component._hf_hook, CpuOffload)
-                    if not is_sequential_cpu_offload:
-                        is_sequential_cpu_offload = (
-                            isinstance(component._hf_hook, AlignDevicesHook)
-                            or hasattr(component._hf_hook, "hooks")
-                            and isinstance(component._hf_hook.hooks[0], AlignDevicesHook)
-                        )
-
-                    logger.info(
-                        "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
-                    )
-                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
-
-        return (is_model_cpu_offload, is_sequential_cpu_offload)
+        return _func_optionally_disable_offloading(_pipeline=_pipeline)
 
     def save_attn_procs(
         self,

src/diffusers/models/transformers/transformer_hunyuan_video.py

@@ -713,14 +713,16 @@ class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin,
         condition_sequence_length = encoder_hidden_states.shape[1]
         sequence_length = latent_sequence_length + condition_sequence_length
         attention_mask = torch.zeros(
-            batch_size, sequence_length, sequence_length, device=hidden_states.device, dtype=torch.bool
-        )  # [B, N, N]
+            batch_size, sequence_length, device=hidden_states.device, dtype=torch.bool
+        )  # [B, N]
 
         effective_condition_sequence_length = encoder_attention_mask.sum(dim=1, dtype=torch.int)  # [B,]
         effective_sequence_length = latent_sequence_length + effective_condition_sequence_length
 
         for i in range(batch_size):
-            attention_mask[i, : effective_sequence_length[i], : effective_sequence_length[i]] = True
+            attention_mask[i, : effective_sequence_length[i]] = True
+        # [B, 1, 1, N], for broadcasting across attention heads
+        attention_mask = attention_mask.unsqueeze(1).unsqueeze(1)
 
         # 4. Transformer blocks
         if torch.is_grad_enabled() and self.gradient_checkpointing:

src/diffusers/utils/__init__.py

@@ -100,7 +100,7 @@ from .import_utils import (
     is_xformers_available,
     requires_backends,
 )
-from .loading_utils import get_module_from_name, load_image, load_video
+from .loading_utils import get_module_from_name, get_submodule_by_name, load_image, load_video
 from .logging import get_logger
 from .outputs import BaseOutput
 from .peft_utils import (

src/diffusers/utils/loading_utils.py

@@ -148,3 +148,15 @@ def get_module_from_name(module, tensor_name: str) -> Tuple[Any, str]:
             module = new_module
         tensor_name = splits[-1]
     return module, tensor_name
+
+
+def get_submodule_by_name(root_module, module_path: str):
+    current = root_module
+    parts = module_path.split(".")
+    for part in parts:
+        if part.isdigit():
+            idx = int(part)
+            current = current[idx]  # e.g., for nn.ModuleList or nn.Sequential
+        else:
+            current = getattr(current, part)
+    return current

tests/lora/test_lora_layers_flux.py

@@ -12,6 +12,7 @@
 # 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 gc
 import os
 import sys
@@ -162,6 +163,105 @@ class FluxLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
         )
         self.assertFalse(np.allclose(images_lora_with_alpha, images_lora, atol=1e-3, rtol=1e-3))
 
+    def test_lora_expansion_works_for_absent_keys(self):
+        components, _, denoiser_lora_config = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        _, _, inputs = self.get_dummy_inputs(with_generator=False)
+
+        output_no_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
+        self.assertTrue(output_no_lora.shape == self.output_shape)
+
+        # Modify the config to have a layer which won't be present in the second LoRA we will load.
+        modified_denoiser_lora_config = copy.deepcopy(denoiser_lora_config)
+        modified_denoiser_lora_config.target_modules.add("x_embedder")
+
+        pipe.transformer.add_adapter(modified_denoiser_lora_config)
+        self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in transformer")
+
+        images_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
+        self.assertFalse(
+            np.allclose(images_lora, output_no_lora, atol=1e-3, rtol=1e-3),
+            "LoRA should lead to different results.",
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            denoiser_state_dict = get_peft_model_state_dict(pipe.transformer)
+            self.pipeline_class.save_lora_weights(tmpdirname, transformer_lora_layers=denoiser_state_dict)
+
+            self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")))
+            pipe.unload_lora_weights()
+            pipe.load_lora_weights(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors"), adapter_name="one")
+
+            # Modify the state dict to exclude "x_embedder" related LoRA params.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors"))
+            lora_state_dict_without_xembedder = {k: v for k, v in lora_state_dict.items() if "x_embedder" not in k}
+
+        pipe.load_lora_weights(lora_state_dict_without_xembedder, adapter_name="two")
+        pipe.set_adapters(["one", "two"])
+        self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in transformer")
+        images_lora_with_absent_keys = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+        self.assertFalse(
+            np.allclose(images_lora, images_lora_with_absent_keys, atol=1e-3, rtol=1e-3),
+            "Different LoRAs should lead to different results.",
+        )
+        self.assertFalse(
+            np.allclose(output_no_lora, images_lora_with_absent_keys, atol=1e-3, rtol=1e-3),
+            "LoRA should lead to different results.",
+        )
+
+    def test_lora_expansion_works_for_extra_keys(self):
+        components, _, denoiser_lora_config = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        _, _, inputs = self.get_dummy_inputs(with_generator=False)
+
+        output_no_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
+        self.assertTrue(output_no_lora.shape == self.output_shape)
+
+        # Modify the config to have a layer which won't be present in the first LoRA we will load.
+        modified_denoiser_lora_config = copy.deepcopy(denoiser_lora_config)
+        modified_denoiser_lora_config.target_modules.add("x_embedder")
+
+        pipe.transformer.add_adapter(modified_denoiser_lora_config)
+        self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in transformer")
+
+        images_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
+        self.assertFalse(
+            np.allclose(images_lora, output_no_lora, atol=1e-3, rtol=1e-3),
+            "LoRA should lead to different results.",
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            denoiser_state_dict = get_peft_model_state_dict(pipe.transformer)
+            self.pipeline_class.save_lora_weights(tmpdirname, transformer_lora_layers=denoiser_state_dict)
+
+            self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")))
+            pipe.unload_lora_weights()
+            # Modify the state dict to exclude "x_embedder" related LoRA params.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors"))
+            lora_state_dict_without_xembedder = {k: v for k, v in lora_state_dict.items() if "x_embedder" not in k}
+            pipe.load_lora_weights(lora_state_dict_without_xembedder, adapter_name="one")
+
+            # Load state dict with `x_embedder`.
+            pipe.load_lora_weights(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors"), adapter_name="two")
+
+        pipe.set_adapters(["one", "two"])
+        self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in transformer")
+        images_lora_with_extra_keys = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+        self.assertFalse(
+            np.allclose(images_lora, images_lora_with_extra_keys, atol=1e-3, rtol=1e-3),
+            "Different LoRAs should lead to different results.",
+        )
+        self.assertFalse(
+            np.allclose(output_no_lora, images_lora_with_extra_keys, atol=1e-3, rtol=1e-3),
+            "LoRA should lead to different results.",
+        )
+
     @unittest.skip("Not supported in Flux.")
     def test_simple_inference_with_text_denoiser_block_scale_for_all_dict_options(self):
         pass
@@ -558,6 +658,131 @@ class FluxControlLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
         self.assertTrue(pipe.transformer.x_embedder.weight.data.shape[1] == in_features * 2)
         self.assertFalse(np.allclose(original_output, lora_output, atol=1e-3, rtol=1e-3))
 
+    def test_lora_unload_with_parameter_expanded_shapes(self):
+        components, _, _ = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
+
+        logger = logging.get_logger("diffusers.loaders.lora_pipeline")
+        logger.setLevel(logging.DEBUG)
+
+        # Change the transformer config to mimic a real use case.
+        num_channels_without_control = 4
+        transformer = FluxTransformer2DModel.from_config(
+            components["transformer"].config, in_channels=num_channels_without_control
+        ).to(torch_device)
+        self.assertTrue(
+            transformer.config.in_channels == num_channels_without_control,
+            f"Expected {num_channels_without_control} channels in the modified transformer but has {transformer.config.in_channels=}",
+        )
+
+        # This should be initialized with a Flux pipeline variant that doesn't accept `control_image`.
+        components["transformer"] = transformer
+        pipe = FluxPipeline(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        _, _, inputs = self.get_dummy_inputs(with_generator=False)
+        control_image = inputs.pop("control_image")
+        original_out = pipe(**inputs, generator=torch.manual_seed(0))[0]
+
+        control_pipe = self.pipeline_class(**components)
+        out_features, in_features = control_pipe.transformer.x_embedder.weight.shape
+        rank = 4
+
+        dummy_lora_A = torch.nn.Linear(2 * in_features, rank, bias=False)
+        dummy_lora_B = torch.nn.Linear(rank, out_features, bias=False)
+        lora_state_dict = {
+            "transformer.x_embedder.lora_A.weight": dummy_lora_A.weight,
+            "transformer.x_embedder.lora_B.weight": dummy_lora_B.weight,
+        }
+        with CaptureLogger(logger) as cap_logger:
+            control_pipe.load_lora_weights(lora_state_dict, "adapter-1")
+            self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
+
+        inputs["control_image"] = control_image
+        lora_out = control_pipe(**inputs, generator=torch.manual_seed(0))[0]
+
+        self.assertFalse(np.allclose(original_out, lora_out, rtol=1e-4, atol=1e-4))
+        self.assertTrue(pipe.transformer.x_embedder.weight.data.shape[1] == 2 * in_features)
+        self.assertTrue(pipe.transformer.config.in_channels == 2 * in_features)
+        self.assertTrue(cap_logger.out.startswith("Expanding the nn.Linear input/output features for module"))
+
+        control_pipe.unload_lora_weights(reset_to_overwritten_params=True)
+        self.assertTrue(
+            control_pipe.transformer.config.in_channels == num_channels_without_control,
+            f"Expected {num_channels_without_control} channels in the modified transformer but has {control_pipe.transformer.config.in_channels=}",
+        )
+        loaded_pipe = FluxPipeline.from_pipe(control_pipe)
+        self.assertTrue(
+            loaded_pipe.transformer.config.in_channels == num_channels_without_control,
+            f"Expected {num_channels_without_control} channels in the modified transformer but has {loaded_pipe.transformer.config.in_channels=}",
+        )
+        inputs.pop("control_image")
+        unloaded_lora_out = loaded_pipe(**inputs, generator=torch.manual_seed(0))[0]
+
+        self.assertFalse(np.allclose(unloaded_lora_out, lora_out, rtol=1e-4, atol=1e-4))
+        self.assertTrue(np.allclose(unloaded_lora_out, original_out, atol=1e-4, rtol=1e-4))
+        self.assertTrue(pipe.transformer.x_embedder.weight.data.shape[1] == in_features)
+        self.assertTrue(pipe.transformer.config.in_channels == in_features)
+
+    def test_lora_unload_with_parameter_expanded_shapes_and_no_reset(self):
+        components, _, _ = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
+
+        logger = logging.get_logger("diffusers.loaders.lora_pipeline")
+        logger.setLevel(logging.DEBUG)
+
+        # Change the transformer config to mimic a real use case.
+        num_channels_without_control = 4
+        transformer = FluxTransformer2DModel.from_config(
+            components["transformer"].config, in_channels=num_channels_without_control
+        ).to(torch_device)
+        self.assertTrue(
+            transformer.config.in_channels == num_channels_without_control,
+            f"Expected {num_channels_without_control} channels in the modified transformer but has {transformer.config.in_channels=}",
+        )
+
+        # This should be initialized with a Flux pipeline variant that doesn't accept `control_image`.
+        components["transformer"] = transformer
+        pipe = FluxPipeline(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        _, _, inputs = self.get_dummy_inputs(with_generator=False)
+        control_image = inputs.pop("control_image")
+        original_out = pipe(**inputs, generator=torch.manual_seed(0))[0]
+
+        control_pipe = self.pipeline_class(**components)
+        out_features, in_features = control_pipe.transformer.x_embedder.weight.shape
+        rank = 4
+
+        dummy_lora_A = torch.nn.Linear(2 * in_features, rank, bias=False)
+        dummy_lora_B = torch.nn.Linear(rank, out_features, bias=False)
+        lora_state_dict = {
+            "transformer.x_embedder.lora_A.weight": dummy_lora_A.weight,
+            "transformer.x_embedder.lora_B.weight": dummy_lora_B.weight,
+        }
+        with CaptureLogger(logger) as cap_logger:
+            control_pipe.load_lora_weights(lora_state_dict, "adapter-1")
+            self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
+
+        inputs["control_image"] = control_image
+        lora_out = control_pipe(**inputs, generator=torch.manual_seed(0))[0]
+
+        self.assertFalse(np.allclose(original_out, lora_out, rtol=1e-4, atol=1e-4))
+        self.assertTrue(pipe.transformer.x_embedder.weight.data.shape[1] == 2 * in_features)
+        self.assertTrue(pipe.transformer.config.in_channels == 2 * in_features)
+        self.assertTrue(cap_logger.out.startswith("Expanding the nn.Linear input/output features for module"))
+
+        control_pipe.unload_lora_weights(reset_to_overwritten_params=False)
+        self.assertTrue(
+            control_pipe.transformer.config.in_channels == 2 * num_channels_without_control,
+            f"Expected {num_channels_without_control} channels in the modified transformer but has {control_pipe.transformer.config.in_channels=}",
+        )
+        no_lora_out = control_pipe(**inputs, generator=torch.manual_seed(0))[0]
+
+        self.assertFalse(np.allclose(no_lora_out, lora_out, rtol=1e-4, atol=1e-4))
+        self.assertTrue(pipe.transformer.x_embedder.weight.data.shape[1] == in_features * 2)
+        self.assertTrue(pipe.transformer.config.in_channels == in_features * 2)
+
     @unittest.skip("Not supported in Flux.")
     def test_simple_inference_with_text_denoiser_block_scale_for_all_dict_options(self):
         pass

tests/lora/test_lora_layers_hunyuanvideo.py

@@ -12,9 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import gc
 import sys
 import unittest
 
+import numpy as np
+import pytest
 import torch
 from transformers import CLIPTextModel, CLIPTokenizer, LlamaModel, LlamaTokenizerFast
 
@@ -26,7 +29,11 @@ from diffusers import (
 )
 from diffusers.utils.testing_utils import (
     floats_tensor,
+    nightly,
+    numpy_cosine_similarity_distance,
+    require_big_gpu_with_torch_cuda,
     require_peft_backend,
+    require_torch_gpu,
     skip_mps,
 )
 
@@ -182,3 +189,69 @@ class HunyuanVideoLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     @unittest.skip("Text encoder LoRA is not supported in HunyuanVideo.")
     def test_simple_inference_with_text_lora_save_load(self):
         pass
+
+
+@nightly
+@require_torch_gpu
+@require_peft_backend
+@require_big_gpu_with_torch_cuda
+@pytest.mark.big_gpu_with_torch_cuda
+class HunyuanVideoLoRAIntegrationTests(unittest.TestCase):
+    """internal note: The integration slices were obtained on DGX.
+
+    torch: 2.5.1+cu124 with CUDA 12.5. Need the same setup for the
+    assertions to pass.
+    """
+
+    num_inference_steps = 10
+    seed = 0
+
+    def setUp(self):
+        super().setUp()
+
+        gc.collect()
+        torch.cuda.empty_cache()
+
+        model_id = "hunyuanvideo-community/HunyuanVideo"
+        transformer = HunyuanVideoTransformer3DModel.from_pretrained(
+            model_id, subfolder="transformer", torch_dtype=torch.bfloat16
+        )
+        self.pipeline = HunyuanVideoPipeline.from_pretrained(
+            model_id, transformer=transformer, torch_dtype=torch.float16
+        ).to("cuda")
+
+    def tearDown(self):
+        super().tearDown()
+
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_original_format_cseti(self):
+        self.pipeline.load_lora_weights(
+            "Cseti/HunyuanVideo-LoRA-Arcane_Jinx-v1", weight_name="csetiarcane-nfjinx-v1-6000.safetensors"
+        )
+        self.pipeline.fuse_lora()
+        self.pipeline.unload_lora_weights()
+        self.pipeline.vae.enable_tiling()
+
+        prompt = "CSETIARCANE. A cat walks on the grass, realistic"
+
+        out = self.pipeline(
+            prompt=prompt,
+            height=320,
+            width=512,
+            num_frames=9,
+            num_inference_steps=self.num_inference_steps,
+            output_type="np",
+            generator=torch.manual_seed(self.seed),
+        ).frames[0]
+        out = out.flatten()
+        out_slice = np.concatenate((out[:8], out[-8:]))
+
+        # fmt: off
+        expected_slice = np.array([0.1013, 0.1924, 0.0078, 0.1021, 0.1929, 0.0078, 0.1023, 0.1919, 0.7402, 0.104, 0.4482, 0.7354, 0.0925, 0.4382, 0.7275, 0.0815])
+        # fmt: on
+
+        max_diff = numpy_cosine_similarity_distance(expected_slice.flatten(), out_slice)
+
+        assert max_diff < 1e-3

tests/quantization/bnb/test_4bit.py

@@ -20,6 +20,7 @@ import unittest
 import numpy as np
 import pytest
 import safetensors.torch
+from huggingface_hub import hf_hub_download
 
 from diffusers import BitsAndBytesConfig, DiffusionPipeline, FluxTransformer2DModel, SD3Transformer2DModel
 from diffusers.utils import is_accelerate_version, logging
@@ -568,6 +569,27 @@ class SlowBnb4BitFluxTests(Base4bitTests):
         max_diff = numpy_cosine_similarity_distance(expected_slice, out_slice)
         self.assertTrue(max_diff < 1e-3)
 
+    def test_lora_loading(self):
+        self.pipeline_4bit.load_lora_weights(
+            hf_hub_download("ByteDance/Hyper-SD", "Hyper-FLUX.1-dev-8steps-lora.safetensors"), adapter_name="hyper-sd"
+        )
+        self.pipeline_4bit.set_adapters("hyper-sd", adapter_weights=0.125)
+
+        output = self.pipeline_4bit(
+            prompt=self.prompt,
+            height=256,
+            width=256,
+            max_sequence_length=64,
+            output_type="np",
+            num_inference_steps=8,
+            generator=torch.Generator().manual_seed(42),
+        ).images
+        out_slice = output[0, -3:, -3:, -1].flatten()
+        expected_slice = np.array([0.5347, 0.5342, 0.5283, 0.5093, 0.4988, 0.5093, 0.5044, 0.5015, 0.4946])
+
+        max_diff = numpy_cosine_similarity_distance(expected_slice, out_slice)
+        self.assertTrue(max_diff < 1e-3)
+
 
 @slow
 class BaseBnb4BitSerializationTests(Base4bitTests):

tests/single_file/test_model_flux_transformer_single_file.py

@@ -0,0 +1,72 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# 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 gc
+import unittest
+
+import torch
+
+from diffusers import (
+    FluxTransformer2DModel,
+)
+from diffusers.utils.testing_utils import (
+    backend_empty_cache,
+    enable_full_determinism,
+    require_torch_accelerator,
+    torch_device,
+)
+
+
+enable_full_determinism()
+
+
+@require_torch_accelerator
+class FluxTransformer2DModelSingleFileTests(unittest.TestCase):
+    model_class = FluxTransformer2DModel
+    ckpt_path = "https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors"
+    alternate_keys_ckpt_paths = ["https://huggingface.co/Comfy-Org/flux1-dev/blob/main/flux1-dev-fp8.safetensors"]
+
+    repo_id = "black-forest-labs/FLUX.1-dev"
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def test_single_file_components(self):
+        model = self.model_class.from_pretrained(self.repo_id, subfolder="transformer")
+        model_single_file = self.model_class.from_single_file(self.ckpt_path)
+
+        PARAMS_TO_IGNORE = ["torch_dtype", "_name_or_path", "_use_default_values", "_diffusers_version"]
+        for param_name, param_value in model_single_file.config.items():
+            if param_name in PARAMS_TO_IGNORE:
+                continue
+            assert (
+                model.config[param_name] == param_value
+            ), f"{param_name} differs between single file loading and pretrained loading"
+
+    def test_checkpoint_loading(self):
+        for ckpt_path in self.alternate_keys_ckpt_paths:
+            torch.cuda.empty_cache()
+            model = self.model_class.from_single_file(ckpt_path)
+
+            del model
+            gc.collect()
+            torch.cuda.empty_cache()