fix Internvl-int8 device map (#937)

Author: hjh0119

docs/source/LLM/命令行参数.md

@@ -123,6 +123,8 @@
 - `--use_loss_scale`: 默认为`False`. 生效时会将Agent的部分字段(Action/Action Input部分)的loss权重加强以强化CoT, 对普通SFT场景没有任何效果.
 - `--custom_register_path`: 默认为`None`. 传入`.py`文件, 用于注册模板、模型和数据集.
 - `--custom_dataset_info`: 默认为`None`, 传入外置dataset_info.json的路径、json字符串或者dict. 用于拓展数据集. 格式参考: https://github.com/modelscope/swift/blob/main/swift/llm/data/dataset_info.json
+- `--device_map_config_path`: 从本地文件中手动配置模型的device_map, 默认为None
+
 
 
 ### FSDP参数

docs/source_en/LLM/Command-line-parameters.md

@@ -123,7 +123,7 @@
 - `--use_loss_scale`: Default is `False`. When taking effect, strengthens loss weight of some Agent fields (Action/Action Input part) to enhance CoT, has no effect in regular SFT scenarios.
 - `--custom_register_path`: Default is `None`. Pass in a `.py` file used to register templates, models, and datasets.
 - `--custom_dataset_info`: Default is `None`. Pass in the path to an external `dataset_info.json`, a JSON string, or a dictionary. Used to register custom datasets. The format example: https://github.com/modelscope/swift/blob/main/swift/llm/data/dataset_info.json
-
+- `device_map_config_path`: Manually configure the model's device map from a local file, defaults to None.
 
 ### FSDP Parameters
 

swift/llm/utils/argument.py

@@ -1019,6 +1019,7 @@ def __post_init__(self) -> None:
             self.load_from_ckpt_dir()
         else:
             assert self.load_dataset_config is False, 'You need to first set `--load_args_from_ckpt_dir true`.'
+        self.handle_compatibility()
         self._handle_dataset_sample()
         self.handle_custom_register()
         self.handle_custom_dataset_info()

swift/llm/utils/model.py

@@ -2609,13 +2609,15 @@ def fix_internvl_inplace_bug(model) -> None:
 
     embedding = model.language_model.get_input_embeddings()
     if not hasattr(embedding, '__old_forward'):  # Avoid double patching
-        if hasattr(embedding, '_old_forward'):  # device_map
-            __old_forward = embedding._old_forward
-            embedding._old_forward = lambda *args, **kwargs: __old_forward(*args, **kwargs).requires_grad_(True).clone()
-        else:
-            __old_forward = embedding.forward
-            embedding.forward = lambda *args, **kwargs: __old_forward(*args, **kwargs).requires_grad_(True).clone()
-        embedding.__old_forward = __old_forward
+        old_forward = embedding.forward
+
+        @wraps(old_forward)
+        def _new_forward(*args, **kwargs):
+            device = args[0].device
+            return old_forward(*args, **kwargs).requires_grad_(True).clone().to(device)
+
+        embedding.__old_forward = old_forward
+        embedding.forward = _new_forward
 
 
 @register_model(