fix #2802

Former-commit-id: 8d8956bad542c0e1c0f7edbf4ffc22bb0f8788ae

src/llmtuner/hparams/model_args.py

@@ -97,6 +97,10 @@ class ModelArguments:
         default=False,
         metadata={"help": "Whether or not to disable CUDA graph in the vLLM engine."},
     )
+    offload_folder: str = field(
+        default="offload",
+        metadata={"help": "Path to offload model weights."},
+    )
     hf_hub_token: Optional[str] = field(
         default=None,
         metadata={"help": "Auth token to log in with Hugging Face Hub."},

src/llmtuner/model/patcher.py

@@ -291,6 +291,9 @@ def patch_config(
         if "device_map" not in init_kwargs:  # quant models cannot use auto device map
             init_kwargs["device_map"] = model_args.device_map or {"": get_current_device()}
 
+        if init_kwargs["device_map"] == "auto":
+            init_kwargs["offload_folder"] = model_args.offload_folder
+
 
 def patch_model(
     model: "PreTrainedModel", tokenizer: "PreTrainedTokenizer", model_args: "ModelArguments", is_trainable: bool