Fixes #12673. `record_stream` in group offloading is not working properly (#12721)

* Fixes #12673.

    Wrong default_stream is used. leading to wrong execution order when record_steram is enabled.

* update

* Update test

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

src/diffusers/hooks/group_offloading.py

@@ -153,27 +153,27 @@ class ModuleGroup:
         finally:
             pinned_dict = None
 
-    def _transfer_tensor_to_device(self, tensor, source_tensor):
+    def _transfer_tensor_to_device(self, tensor, source_tensor, default_stream):
         tensor.data = source_tensor.to(self.onload_device, non_blocking=self.non_blocking)
         if self.record_stream:
-            tensor.data.record_stream(self._torch_accelerator_module.current_stream())
+            tensor.data.record_stream(default_stream)
 
-    def _process_tensors_from_modules(self, pinned_memory=None):
+    def _process_tensors_from_modules(self, pinned_memory=None, default_stream=None):
         for group_module in self.modules:
             for param in group_module.parameters():
                 source = pinned_memory[param] if pinned_memory else param.data
-                self._transfer_tensor_to_device(param, source)
+                self._transfer_tensor_to_device(param, source, default_stream)
             for buffer in group_module.buffers():
                 source = pinned_memory[buffer] if pinned_memory else buffer.data
-                self._transfer_tensor_to_device(buffer, source)
+                self._transfer_tensor_to_device(buffer, source, default_stream)
 
         for param in self.parameters:
             source = pinned_memory[param] if pinned_memory else param.data
-            self._transfer_tensor_to_device(param, source)
+            self._transfer_tensor_to_device(param, source, default_stream)
 
         for buffer in self.buffers:
             source = pinned_memory[buffer] if pinned_memory else buffer.data
-            self._transfer_tensor_to_device(buffer, source)
+            self._transfer_tensor_to_device(buffer, source, default_stream)
 
     def _onload_from_disk(self):
         if self.stream is not None:
@@ -208,10 +208,12 @@ class ModuleGroup:
             self.stream.synchronize()
 
         context = nullcontext() if self.stream is None else self._torch_accelerator_module.stream(self.stream)
+        default_stream = self._torch_accelerator_module.current_stream() if self.stream is not None else None
+
         with context:
             if self.stream is not None:
                 with self._pinned_memory_tensors() as pinned_memory:
-                    self._process_tensors_from_modules(pinned_memory)
+                    self._process_tensors_from_modules(pinned_memory, default_stream=default_stream)
             else:
                 self._process_tensors_from_modules(None)
 

tests/models/test_modeling_common.py

@@ -1814,9 +1814,6 @@ class ModelTesterMixin:
             torch.manual_seed(0)
             return model(**inputs_dict)[0]
 
-        if self.__class__.__name__ == "AutoencoderKLCosmosTests" and offload_type == "leaf_level":
-            pytest.skip("With `leaf_type` as the offloading type, it fails. Needs investigation.")
-
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
         torch.manual_seed(0)
         model = self.model_class(**init_dict)