Merge branch 'main' into specific-accelerate-imports

src/diffusers/commands/env.py

@@ -83,9 +83,9 @@ class EnvironmentCommand(BaseDiffusersCLICommand):
 
         accelerate_version = "not installed"
         if is_accelerate_available():
-            import accelerate
+            from accelerate import __version__
 
-            accelerate_version = accelerate.__version__
+            accelerate_version = __version__
 
         peft_version = "not installed"
         if is_peft_available():

src/diffusers/models/modeling_utils.py

@@ -72,7 +72,7 @@ else:
 
 
 if is_accelerate_available():
-    import accelerate
+    from accelerate import init_empty_weights, load_checkpoint_and_dispatch
 
 
 def get_parameter_device(parameter: torch.nn.Module) -> torch.device:
@@ -736,7 +736,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
 
             if low_cpu_mem_usage:
                 # Instantiate model with empty weights
-                with accelerate.init_empty_weights():
+                with init_empty_weights():
                     model = cls.from_config(config, **unused_kwargs)
 
                 # if device_map is None, load the state dict and move the params from meta device to the cpu
@@ -781,7 +781,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                         device_map = {"": "cpu"}
                         force_hook = False
                     try:
-                        accelerate.load_checkpoint_and_dispatch(
+                        load_checkpoint_and_dispatch(
                             model,
                             model_file if not is_sharded else sharded_ckpt_cached_folder,
                             device_map,
@@ -811,7 +811,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                                 " please also re-upload it or open a PR on the original repository."
                             )
                             model._temp_convert_self_to_deprecated_attention_blocks()
-                            accelerate.load_checkpoint_and_dispatch(
+                            load_checkpoint_and_dispatch(
                                 model,
                                 model_file if not is_sharded else sharded_ckpt_cached_folder,
                                 device_map,

src/diffusers/pipelines/pipeline_loading_utils.py

@@ -50,8 +50,7 @@ if is_transformers_available():
     from transformers.utils import WEIGHTS_NAME as TRANSFORMERS_WEIGHTS_NAME
 
 if is_accelerate_available():
-    import accelerate
-    from accelerate import dispatch_model
+    from accelerate import dispatch_model, init_empty_weights
     from accelerate.hooks import remove_hook_from_module
     from accelerate.utils import compute_module_sizes, get_max_memory
 
@@ -443,7 +442,7 @@ def _load_empty_model(
             subfolder=kwargs.pop("subfolder", None),
             user_agent=user_agent,
         )
-        with accelerate.init_empty_weights():
+        with init_empty_weights():
             model = class_obj.from_config(config, **unused_kwargs)
     elif is_transformers_model:
         config_class = getattr(class_obj, "config_class", None)
@@ -461,7 +460,7 @@ def _load_empty_model(
             revision=kwargs.pop("revision", None),
             user_agent=user_agent,
         )
-        with accelerate.init_empty_weights():
+        with init_empty_weights():
             model = class_obj(config)
 
     if model is not None:
@@ -529,7 +528,7 @@ def _get_final_device_map(device_map, pipeline_class, passed_class_obj, init_dic
                 name,
                 is_pipeline_module,
             )
-            with accelerate.init_empty_weights():
+            with init_empty_weights():
                 loaded_sub_model = passed_class_obj[name]
 
         else:

src/diffusers/pipelines/pipeline_utils.py

@@ -85,7 +85,7 @@ from .pipeline_loading_utils import (
 
 
 if is_accelerate_available():
-    import accelerate
+    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
 
 
 LIBRARIES = []
@@ -377,16 +377,16 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 return False
 
             return hasattr(module, "_hf_hook") and (
-                isinstance(module._hf_hook, accelerate.hooks.AlignDevicesHook)
+                isinstance(module._hf_hook, AlignDevicesHook)
                 or hasattr(module._hf_hook, "hooks")
-                and isinstance(module._hf_hook.hooks[0], accelerate.hooks.AlignDevicesHook)
+                and isinstance(module._hf_hook.hooks[0], AlignDevicesHook)
             )
 
         def module_is_offloaded(module):
             if not is_accelerate_available() or is_accelerate_version("<", "0.17.0.dev0"):
                 return False
 
-            return hasattr(module, "_hf_hook") and isinstance(module._hf_hook, accelerate.hooks.CpuOffload)
+            return hasattr(module, "_hf_hook") and isinstance(module._hf_hook, CpuOffload)
 
         # .to("cuda") would raise an error if the pipeline is sequentially offloaded, so we raise our own to make it clearer
         pipeline_is_sequentially_offloaded = any(
@@ -1009,7 +1009,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         """
         for _, model in self.components.items():
             if isinstance(model, torch.nn.Module) and hasattr(model, "_hf_hook"):
-                accelerate.hooks.remove_hook_from_module(model, recurse=True)
+                remove_hook_from_module(model, recurse=True)
         self._all_hooks = []
 
     def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):

src/diffusers/utils/accelerate_utils.py

@@ -21,7 +21,7 @@ from .import_utils import is_accelerate_available
 
 
 if is_accelerate_available():
-    import accelerate
+    from accelerate import __version__
 
 
 def apply_forward_hook(method):
@@ -36,7 +36,7 @@ def apply_forward_hook(method):
     """
     if not is_accelerate_available():
         return method
-    accelerate_version = version.parse(accelerate.__version__).base_version
+    accelerate_version = version.parse(__version__).base_version
     if version.parse(accelerate_version) < version.parse("0.17.0"):
         return method