Fix offload disk for loading derivated model checkpoint into base model (#27253)

* fix

* style

* add test

src/transformers/modeling_utils.py

@@ -3793,8 +3793,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         else:
             folder = None
         if device_map is not None and is_safetensors:
-            param_device_map = expand_device_map(device_map, original_loaded_keys)
+            param_device_map = expand_device_map(device_map, original_loaded_keys, start_prefix)
-
             str_dtype = str(dtype).replace("torch.", "") if dtype is not None else "float32"
             if sharded_metadata is None:
                 archive_file = (
@@ -3806,9 +3805,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             else:
                 weight_map = {p: os.path.join(folder, f) for p, f in sharded_metadata["weight_map"].items()}
             offload_index = {
-                p: {"safetensors_file": f, "weight_name": p, "dtype": str_dtype}
+                p[len(start_prefix) :]: {"safetensors_file": f, "weight_name": p, "dtype": str_dtype}
                 for p, f in weight_map.items()
-                if param_device_map[p] == "disk"
+                if p.startswith(start_prefix) and param_device_map[p[len(start_prefix) :]] == "disk"
             }
 
         if state_dict is not None:
@@ -3842,7 +3841,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 state_dict_index = None
 
             if is_sharded_safetensors:
-                disk_only_shard_files = get_disk_only_shard_files(device_map, sharded_metadata=sharded_metadata)
+                disk_only_shard_files = get_disk_only_shard_files(
+                    device_map, sharded_metadata=sharded_metadata, start_prefix=start_prefix
+                )
                 disk_only_shard_files = [os.path.join(folder, f) for f in disk_only_shard_files]
             else:
                 disk_only_shard_files = []
@@ -4576,11 +4577,12 @@ def unwrap_model(model: nn.Module) -> nn.Module:
         return model
 
 
-def expand_device_map(device_map, param_names):
+def expand_device_map(device_map, param_names, start_prefix):
     """
     Expand a device map to return the correspondance parameter name to device.
     """
     new_device_map = {}
+    param_names = [p[len(start_prefix) :] for p in param_names if p.startswith(start_prefix)]
     for module, device in device_map.items():
         new_device_map.update(
             {p: device for p in param_names if p == module or p.startswith(f"{module}.") or module == ""}
@@ -4588,12 +4590,16 @@ def expand_device_map(device_map, param_names):
     return new_device_map
 
 
-def get_disk_only_shard_files(device_map, sharded_metadata):
+def get_disk_only_shard_files(device_map, sharded_metadata, start_prefix):
     """
     Returns the list of shard files containing only weights offloaded to disk.
     """
+
+    weight_map = {
+        p[len(start_prefix) :]: v for p, v in sharded_metadata["weight_map"].items() if p.startswith(start_prefix)
+    }
     files_content = collections.defaultdict(list)
-    for weight_name, filename in sharded_metadata["weight_map"].items():
+    for weight_name, filename in weight_map.items():
         while len(weight_name) > 0 and weight_name not in device_map:
             weight_name = ".".join(weight_name.split(".")[:-1])
         files_content[filename].append(device_map[weight_name])

tests/test_modeling_utils.py

@@ -750,6 +750,46 @@ class ModelUtilsTest(TestCasePlus):
 
             self.assertTrue(torch.allclose(outputs1.logits.cpu(), outputs2.logits.cpu()))
 
+    @require_accelerate
+    @mark.accelerate_tests
+    @require_torch_accelerator
+    def test_from_pretrained_disk_offload_derived_to_base_model(self):
+        derived_model = AutoModelForCausalLM.from_pretrained("hf-internal-testing/tiny-random-gpt2")
+
+        device_map = {
+            "wte": 0,
+            "wpe": 0,
+            "h.0": "cpu",
+            "h.1": "cpu",
+            "h.2": "cpu",
+            "h.3": "disk",
+            "h.4": "disk",
+            "ln_f": 0,
+        }
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            inputs = torch.tensor([[1, 2, 3]]).to(0)
+            derived_model.save_pretrained(tmp_dir, use_safetensors=True)
+            base_model = AutoModel.from_pretrained(tmp_dir)
+            outputs1 = base_model.to(0)(inputs)
+
+            # with disk offload
+            offload_folder = os.path.join(tmp_dir, "offload")
+            base_model_with_offload = AutoModel.from_pretrained(
+                tmp_dir, device_map=device_map, offload_folder=offload_folder
+            )
+            outputs2 = base_model_with_offload(inputs)
+            self.assertTrue(torch.allclose(outputs1[0].cpu(), outputs2[0].cpu()))
+
+            # With state dict temp offload
+            new_model_with_offload = AutoModel.from_pretrained(
+                tmp_dir,
+                device_map=device_map,
+                offload_folder=offload_folder,
+                offload_state_dict=True,
+            )
+            outputs2 = new_model_with_offload(inputs)
+            self.assertTrue(torch.allclose(outputs1[0].cpu(), outputs2[0].cpu()))
+
     def test_cached_files_are_used_when_internet_is_down(self):
         # A mock response for an HTTP head request to emulate server down
         response_mock = mock.Mock()