[FIX] `offload_weight()` takes from 3 to 4 positional arguments but 5 were given (#29457)

* use require_torch_gpu

* enable on XPU

* fix

src/transformers/modeling_utils.py

@@ -796,7 +796,7 @@ def _load_state_dict_into_meta_model(
             if not is_safetensors:
                 offload_index = offload_weight(param, param_name, offload_folder, offload_index)
         elif param_device == "cpu" and state_dict_index is not None:
-            state_dict_index = offload_weight(param, param_name, model, state_dict_folder, state_dict_index)
+            state_dict_index = offload_weight(param, param_name, state_dict_folder, state_dict_index)
         elif (
             hf_quantizer is None
             or (not hf_quantizer.requires_parameters_quantization)

tests/test_modeling_utils.py

@@ -765,7 +765,7 @@ class ModelUtilsTest(TestCasePlus):
 
     @require_accelerate
     @mark.accelerate_tests
-    @require_torch_accelerator
+    @require_torch_gpu
     def test_from_pretrained_disk_offload_task_model(self):
         model = AutoModel.from_pretrained("hf-internal-testing/tiny-random-gpt2")
         device_map = {
@@ -808,7 +808,7 @@ class ModelUtilsTest(TestCasePlus):
 
     @require_accelerate
     @mark.accelerate_tests
-    @require_torch_accelerator
+    @require_torch_gpu
     def test_from_pretrained_disk_offload_derived_to_base_model(self):
         derived_model = AutoModelForCausalLM.from_pretrained("hf-internal-testing/tiny-random-gpt2")