Enable modelscope for itrex

intel_extension_for_transformers/transformers/modeling/modeling_auto.py

@@ -588,6 +588,7 @@ def forward(self, input: torch.Tensor) -> tuple[torch.Tensor, None]:
                         pretrained_model_name_or_path,
                         *model_args,
                         config=config,
+                        model_hub=model_hub,
                         **kwargs,
                     )
                     logger.info(
@@ -1452,9 +1453,13 @@ def train_func(model):
                 model.quantization_config = None
                 return model
             else:
-                model = cls.ORIG_MODEL.from_pretrained(
-                    pretrained_model_name_or_path, *model_args, config=config, **kwargs
-                )
+                if model_hub=="modelscope":
+                    from modelscope import AutoModelForCausalLM # pylint: disable=E0401
+                    model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path,trust_remote_code=True)
+                else:
+                    model = cls.ORIG_MODEL.from_pretrained(
+                        pretrained_model_name_or_path, *model_args, config=config, **kwargs
+                    )
                 if (
                     not torch.cuda.is_available()
                     or device_map == "cpu"
@@ -1520,7 +1525,7 @@ def load_low_bit(cls, pretrained_model_name_or_path, *model_args, **kwargs):
         device_map = kwargs.pop("device_map", "auto")
         use_safetensors = kwargs.pop("use_safetensors", None)
         kwarg_attn_imp = kwargs.pop("attn_implementation", None)
-
+        model_hub = kwargs.pop("model_hub", None)
         # lm-eval device map is dictionary
         device_map = device_map[""] if isinstance(device_map, dict) and "" in device_map else device_map
 
@@ -1709,7 +1714,20 @@ def load_low_bit(cls, pretrained_model_name_or_path, *model_args, **kwargs):
                         "_raise_exceptions_for_missing_entries": False,
                         "_commit_hash": commit_hash,
                     }
-                    resolved_archive_file = cached_file(pretrained_model_name_or_path, filename, **cached_file_kwargs)
+                    if model_hub=="modelscope":
+                        from modelscope import snapshot_download # pylint: disable=E0401
+                        model_dir = snapshot_download(pretrained_model_name_or_path)
+                        if os.path.exists(model_dir+"/model.safetensors"):
+                            resolved_archive_file = model_dir+"/model.safetensors"
+                        elif os.path.exists(model_dir+"/pytorch_model.bin"):
+                            resolved_archive_file = model_dir+"/pytorch_model.bin"
+                        else:
+                            assert (
+                                    resolved_archive_file  is not None
+                                    ), "Don't detect this model checkpoint"
+                    else:
+                        resolved_archive_file = cached_file(pretrained_model_name_or_path, filename,
+                                                            **cached_file_kwargs)
 
                     # Since we set _raise_exceptions_for_missing_entries=False, we don't get an exception but a None
                     # result when internet is up, the repo and revision exist, but the file does not.