update

Author: a-r-r-o-w

src/diffusers/loaders/single_file_model.py

@@ -431,10 +431,7 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
                 keep_in_fp32_modules=keep_in_fp32_modules,
                 unexpected_keys=unexpected_keys,
             )
-            # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
-            # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
             empty_device_cache()
-            device_synchronize()
         else:
             _, unexpected_keys = model.load_state_dict(diffusers_format_checkpoint, strict=False)
 

src/diffusers/loaders/single_file_utils.py

@@ -1690,10 +1690,7 @@ def create_diffusers_clip_model_from_ldm(
 
     if is_accelerate_available():
         load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
-        # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
-        # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
         empty_device_cache()
-        device_synchronize()
     else:
         model.load_state_dict(diffusers_format_checkpoint, strict=False)
 
@@ -2153,10 +2150,7 @@ def create_diffusers_t5_model_from_checkpoint(
 
     if is_accelerate_available():
         load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
-        # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
-        # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
         empty_device_cache()
-        device_synchronize()
     else:
         model.load_state_dict(diffusers_format_checkpoint)
 

src/diffusers/loaders/transformer_flux.py

@@ -82,7 +82,6 @@ def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict, low_cpu_mem_us
             device_map = {"": self.device}
             load_model_dict_into_meta(image_projection, updated_state_dict, device_map=device_map, dtype=self.dtype)
             empty_device_cache()
-            device_synchronize()
 
         return image_projection
 
@@ -158,7 +157,6 @@ def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=_
                 key_id += 1
 
         empty_device_cache()
-        device_synchronize()
 
         return attn_procs
 

src/diffusers/loaders/transformer_sd3.py

@@ -82,7 +82,6 @@ def _convert_ip_adapter_attn_to_diffusers(
                 )
 
         empty_device_cache()
-        device_synchronize()
 
         return attn_procs
 
@@ -152,7 +151,6 @@ def _convert_ip_adapter_image_proj_to_diffusers(
             device_map = {"": self.device}
             load_model_dict_into_meta(image_proj, updated_state_dict, device_map=device_map, dtype=self.dtype)
             empty_device_cache()
-            device_synchronize()
 
         return image_proj
 

src/diffusers/loaders/unet.py

@@ -755,7 +755,6 @@ def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict, low_cpu_mem_us
             device_map = {"": self.device}
             load_model_dict_into_meta(image_projection, updated_state_dict, device_map=device_map, dtype=self.dtype)
             empty_device_cache()
-            device_synchronize()
 
         return image_projection
 
@@ -854,7 +853,6 @@ def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=_
                 key_id += 2
 
         empty_device_cache()
-        device_synchronize()
 
         return attn_procs
 

src/diffusers/models/modeling_utils.py

@@ -1540,10 +1540,7 @@ def _load_pretrained_model(
                     assign_to_params_buffers = check_support_param_buffer_assignment(model, state_dict)
                 error_msgs += _load_state_dict_into_model(model, state_dict, assign_to_params_buffers)
 
-        # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
-        # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
         empty_device_cache()
-        device_synchronize()
 
         if offload_index is not None and len(offload_index) > 0:
             save_offload_index(offload_index, offload_folder)