fix #2735

Former-commit-id: 416f6333f66b6afd70a3a936d82593efca583235

src/llmtuner/model/loader.py

@@ -5,7 +5,7 @@ from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
 from trl import AutoModelForCausalLMWithValueHead
 
 from ..extras.logging import get_logger
-from ..extras.misc import check_dependencies, count_parameters, get_current_device, try_download_model_from_ms
+from ..extras.misc import count_parameters, get_current_device, try_download_model_from_ms
 from .adapter import init_adapter
 from .patcher import patch_config, patch_model, patch_tokenizer, patch_valuehead_model
 from .utils import load_valuehead_params, register_autoclass
@@ -20,9 +20,6 @@ if TYPE_CHECKING:
 logger = get_logger(__name__)
 
 
-check_dependencies()
-
-
 def _get_init_kwargs(model_args: "ModelArguments") -> Dict[str, Any]:
     return {
         "trust_remote_code": True,

src/llmtuner/model/patcher.py

@@ -286,9 +286,9 @@ def patch_config(
 
     init_kwargs["torch_dtype"] = model_args.compute_dtype
     if not is_deepspeed_zero3_enabled():
-        init_kwargs["low_cpu_mem_usage"] = True
-        if "device_map" not in init_kwargs:
-            init_kwargs["device_map"] = {"": get_current_device()} if is_trainable else "auto"
+        init_kwargs["low_cpu_mem_usage"] = model_args.low_cpu_mem_usage
+        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()}
 
 
 def patch_model(