fix argument (#1066)

Author: hjh0119

swift/llm/sft.py

@@ -54,7 +54,7 @@ def llm_sft(args: SftArguments) -> Dict[str, Union[str, Any]]:
         config_path = args.device_map_config_path if os.path.isabs(args.device_map_config_path) else os.path.join(
             cwd, args.device_map_config_path)
         with open(config_path, 'r') as json_file:
-            model_kwargs['device_map'] = json.load(json_file)
+            model_kwargs = {'device_map': json.load(json_file)}
     else:
         model_kwargs = {'low_cpu_mem_usage': True}
         if is_dist() and not is_ddp_plus_mp():

swift/llm/utils/argument.py

@@ -152,6 +152,10 @@ def select_bnb(self: Union['SftArguments', 'InferArguments']) -> Tuple[Optional[
             elif quantization_bit == 8:
                 require_version('bitsandbytes')
                 load_in_4bit, load_in_8bit = False, True
+            else:
+                logger.warning('bnb only support 4/8 bits quantization, you should assign --quantization_bit 4 or 8,\
+                    Or specify another quantization method; No quantization will be performed here.')
+                load_in_4bit, load_in_8bit = False, False
         else:
             load_in_4bit, load_in_8bit = False, False
 

swift/llm/utils/model.py

@@ -14,7 +14,6 @@
 import transformers
 from modelscope import (AutoConfig, AutoModel, AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig,
                         GenerationConfig, GPTQConfig, snapshot_download)
-from modelscope.hub.utils.utils import get_cache_dir
 from packaging import version
 from torch import Tensor
 from torch import dtype as Dtype
@@ -25,6 +24,7 @@
 from transformers.utils.versions import require_version
 
 from swift import get_logger
+from swift.hub.utils.utils import get_cache_dir
 from swift.utils import get_dist_setting, safe_ddp_context, subprocess_run, use_torchacc
 from .template import TemplateType
 from .utils import get_max_model_len, is_unsloth_available