fix bug for _load_pretrained_model on windows

src/transformers/modeling_utils.py

@@ -4215,14 +4215,31 @@ def _load_pretrained_model(
                 pattern = re.compile(r"(" + "|".join(map(re.escape, keys)) + r")")
                 if sharded_metadata is None:
                     k_v_iterator = dict.fromkeys(
-                        safe_open(checkpoint_files[0], framework="pt").keys(), checkpoint_files[0].rsplit("/", 1)[1]
+                        safe_open(checkpoint_files[0], framework="pt").keys(), os.path.basename(checkpoint_files[0])
                     ).items()
                 else:
                     k_v_iterator = sharded_metadata["weight_map"].items()
 
-                merged_state_dict = {}
+                # Create a mapping from filename to full path for all checkpoint files
+                filename_to_path = {os.path.basename(f): f for f in checkpoint_files}
+
+                # Group weights by file to load sequentially and avoid keeping too many files open
+                weights_by_file = {}
                 for k, v in k_v_iterator:
-                    match = pattern.match(k)
+                    if v not in weights_by_file:
+                        weights_by_file[v] = []
+                    weights_by_file[v].append(k)
+
+                merged_state_dict = {}
+                # Load each file sequentially
+                for filename, weight_keys in weights_by_file.items():
+                    # Use the mapping to get the correct file path instead of joining paths
+                    # This handles symbolic links on Windows correctly
+                    shard_file_path = filename_to_path.get(
+                        filename, os.path.join(os.path.dirname(checkpoint_files[0]), filename)
+                    )
+
+                    match = pattern.match(weight_keys[0])
                     if match and match.group(1) != "":
                         device = device_map[match.group(1)]
                     else:
@@ -4231,11 +4248,11 @@ def _load_pretrained_model(
                             device = device.index  # safetensors only
                     if device == "disk":
                         device = "cpu"  # we read to cpu to then write to disk
-                    file_pointer = safe_open(
-                        os.path.join(checkpoint_files[0].rsplit("/", 1)[0], v), framework="pt", device=device
-                    )
-                    all_pointer.add(file_pointer)
-                    merged_state_dict[k] = file_pointer.get_slice(k)  # don't materialize yet
+
+                    with safe_open(shard_file_path, framework="pt", device=device) as f:
+                        for k in weight_keys:
+                            # Materialize the tensor immediately instead of keeping a lazy slice
+                            merged_state_dict[k] = f.get_tensor(k)
             elif state_dict is not None:
                 merged_state_dict = state_dict
             elif checkpoint_files is not None: