[FEAT] Model loading refactor

src/diffusers/loaders/single_file_model.py

@@ -52,7 +52,7 @@
 
 
 if is_accelerate_available():
-    from accelerate import init_empty_weights
+    from accelerate import dispatch_model, init_empty_weights
 
     from ..models.modeling_utils import load_model_dict_into_meta
 
@@ -366,19 +366,23 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
                 keep_in_fp32_modules=keep_in_fp32_modules,
             )
 
+        device_map = None
         if is_accelerate_available():
             param_device = torch.device(device) if device else torch.device("cpu")
-            named_buffers = model.named_buffers()
-            unexpected_keys = load_model_dict_into_meta(
+            empty_state_dict = model.state_dict()
+            unexpected_keys = [
+                param_name for param_name in diffusers_format_checkpoint if param_name not in empty_state_dict
+            ]
+            device_map = {"": param_device}
+            load_model_dict_into_meta(
                 model,
                 diffusers_format_checkpoint,
                 dtype=torch_dtype,
-                device=param_device,
+                device_map=device_map,
                 hf_quantizer=hf_quantizer,
                 keep_in_fp32_modules=keep_in_fp32_modules,
-                named_buffers=named_buffers,
+                unexpected_keys=unexpected_keys,
             )
-
         else:
             _, unexpected_keys = model.load_state_dict(diffusers_format_checkpoint, strict=False)
 
@@ -400,4 +404,8 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
 
         model.eval()
 
+        if device_map is not None:
+            device_map_kwargs = {"device_map": device_map}
+            dispatch_model(model, **device_map_kwargs)
+
         return model

src/diffusers/loaders/single_file_utils.py

@@ -1593,18 +1593,9 @@ def create_diffusers_clip_model_from_ldm(
         raise ValueError("The provided checkpoint does not seem to contain a valid CLIP model.")
 
     if is_accelerate_available():
-        unexpected_keys = load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
+        load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
     else:
-        _, unexpected_keys = model.load_state_dict(diffusers_format_checkpoint, strict=False)
-
-    if model._keys_to_ignore_on_load_unexpected is not None:
-        for pat in model._keys_to_ignore_on_load_unexpected:
-            unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
-
-    if len(unexpected_keys) > 0:
-        logger.warning(
-            f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
-        )
+        model.load_state_dict(diffusers_format_checkpoint, strict=False)
 
     if torch_dtype is not None:
         model.to(torch_dtype)
@@ -2061,16 +2052,7 @@ def create_diffusers_t5_model_from_checkpoint(
     diffusers_format_checkpoint = convert_sd3_t5_checkpoint_to_diffusers(checkpoint)
 
     if is_accelerate_available():
-        unexpected_keys = load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
-        if model._keys_to_ignore_on_load_unexpected is not None:
-            for pat in model._keys_to_ignore_on_load_unexpected:
-                unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
-
-        if len(unexpected_keys) > 0:
-            logger.warning(
-                f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
-            )
-
+        load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
     else:
         model.load_state_dict(diffusers_format_checkpoint)