[LoRA] feat: support loading loras into 4bit quantized Flux models.

src/diffusers/loaders/lora_pipeline.py

@@ -21,6 +21,7 @@
 from ..utils import (
     USE_PEFT_BACKEND,
     deprecate,
+    get_submodule_by_name,
     is_peft_available,
     is_peft_version,
     is_torch_version,
@@ -1981,10 +1982,17 @@ def _maybe_expand_transformer_param_shape_or_error_(
                 in_features = state_dict[lora_A_weight_name].shape[1]
                 out_features = state_dict[lora_B_weight_name].shape[0]
 
+                # Model maybe loaded with different quantization schemes which may flatten the params.
+                # `bitsandbytes`, for example, flatten the weights when using 4bit. 8bit bnb models
+                # preserve weight shape.
+                module_weight_shape = cls._calculate_module_shape(model=transformer, base_module=module)
+
                 # This means there's no need for an expansion in the params, so we simply skip.
-                if tuple(module_weight.shape) == (out_features, in_features):
+                if tuple(module_weight_shape) == (out_features, in_features):
                     continue
 
+                # TODO (sayakpaul): We still need to consider if the module we're expanding is
+                # quantized and handle it accordingly if that is the case.
                 module_out_features, module_in_features = module_weight.shape
                 debug_message = ""
                 if in_features > module_in_features:
@@ -2080,13 +2088,16 @@ def _maybe_expand_lora_state_dict(cls, transformer, lora_state_dict):
             base_weight_param = transformer_state_dict[base_param_name]
             lora_A_param = lora_state_dict[f"{prefix}{k}.lora_A.weight"]
 
-            if base_weight_param.shape[1] > lora_A_param.shape[1]:
+            # TODO (sayakpaul): Handle the cases when we actually need to expand when using quantization.
+            base_module_shape = cls._calculate_module_shape(model=transformer, base_weight_param_name=base_param_name)
+
+            if base_module_shape[1] > lora_A_param.shape[1]:
                 shape = (lora_A_param.shape[0], base_weight_param.shape[1])
                 expanded_state_dict_weight = torch.zeros(shape, device=base_weight_param.device)
                 expanded_state_dict_weight[:, : lora_A_param.shape[1]].copy_(lora_A_param)
                 lora_state_dict[f"{prefix}{k}.lora_A.weight"] = expanded_state_dict_weight
                 expanded_module_names.add(k)
-            elif base_weight_param.shape[1] < lora_A_param.shape[1]:
+            elif base_module_shape[1] < lora_A_param.shape[1]:
                 raise NotImplementedError(
                     f"This LoRA param ({k}.lora_A.weight) has an incompatible shape {lora_A_param.shape}. Please open an issue to file for a feature request - https://github.com/huggingface/diffusers/issues/new."
                 )
@@ -2098,6 +2109,28 @@ def _maybe_expand_lora_state_dict(cls, transformer, lora_state_dict):
 
         return lora_state_dict
 
+    @staticmethod
+    def _calculate_module_shape(
+        model: "torch.nn.Module",
+        base_module: "torch.nn.Linear" = None,
+        base_weight_param_name: str = None,
+    ) -> "torch.Size":
+        def _get_weight_shape(weight: torch.Tensor):
+            return weight.quant_state.shape if weight.__class__.__name__ == "Params4bit" else weight.shape
+
+        if base_module is not None:
+            return _get_weight_shape(base_module.weight)
+        elif base_weight_param_name is not None:
+            if not base_weight_param_name.endswith(".weight"):
+                raise ValueError(
+                    f"Invalid `base_weight_param_name` passed as it does not end with '.weight' {base_weight_param_name=}."
+                )
+            module_path = base_weight_param_name.rsplit(".weight", 1)[0]
+            submodule = get_submodule_by_name(model, module_path)
+            return _get_weight_shape(submodule.weight)
+
+        raise ValueError("Either `base_module` or `base_weight_param_name` must be provided.")
+
 
 # The reason why we subclass from `StableDiffusionLoraLoaderMixin` here is because Amused initially
 # relied on `StableDiffusionLoraLoaderMixin` for its LoRA support.

src/diffusers/utils/__init__.py

@@ -101,7 +101,7 @@
     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/quantization/bnb/test_4bit.py

@@ -20,6 +20,7 @@
 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 @@ def test_quality(self):
         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):