feat: load model from modelscope

README.md

@@ -161,6 +161,17 @@ result = model.generate("Uncovering deep insights begins with")[0] # tokens
 print(model.tokenizer.decode(result)) # string output
 ```
 
+To use models from [ModelScope](https://www.modelscope.cn/) instead of HuggingFace Hub, set an environment variable:
+```shell
+export GPTQMODEL_USE_MODELSCOPE=True
+```
+```py
+from gptqmodel import GPTQModel
+# load Qwen/Qwen2.5-0.5B-Instruct-GPTQ-Int4 from modelscope
+model = GPTQModel.load("Qwen/Qwen2.5-0.5B-Instruct-GPTQ-Int4")
+result = model.generate("Uncovering deep insights begins with")[0] # tokens
+print(model.tokenizer.decode(result)) # string output
+```
 
 ### OpenAI API compatible end-point
 ```py

gptqmodel/__init__.py

@@ -19,3 +19,8 @@
 from .utils import BACKEND
 from .utils.exllama import exllama_set_max_input_length
 from .version import __version__
+
+import os
+if os.getenv('GPTQMODEL_USE_MODELSCOPE', 'False').lower() in ['true', '1']:
+    from modelscope.utils.hf_util.patcher import patch_hub
+    patch_hub()

gptqmodel/models/loader.py

@@ -23,7 +23,10 @@
 
 import torch
 import transformers
-from huggingface_hub import snapshot_download
+if os.getenv('GPTQMODEL_USE_MODELSCOPE', 'False').lower() in ['true', '1']:
+    from modelscope import snapshot_download
+else:
+    from huggingface_hub import snapshot_download
 from packaging.version import InvalidVersion, Version
 from transformers import AutoConfig, AutoTokenizer, PretrainedConfig
 from transformers.modeling_utils import no_init_weights

tests/test_modelscope.py

@@ -0,0 +1,27 @@
+import os
+os.environ["GPTQMODEL_USE_MODELSCOPE"] = "True"
+import sys
+import subprocess  # noqa: E402
+import importlib.util  # noqa: E402
+
+if importlib.util.find_spec("modelscope") is None:
+    subprocess.check_call([sys.executable, "-m", "pip", "install", "modelscope", "-U"])
+
+from models.model_test import ModelTest  # noqa: E402
+from gptqmodel import GPTQModel
+
+
+class TestLoadModelscope(ModelTest):
+
+    @classmethod
+    def setUpClass(self):
+        self.MODEL_ID = "Qwen/Qwen2.5-0.5B-Instruct-GPTQ-Int4"
+
+    def test_load_modelscope(self):
+        model = GPTQModel.load(self.MODEL_ID)
+
+        result = model.generate("The capital of mainland China is")[0]
+        str_output = model.tokenizer.decode(result)
+        assert "beijing" in str_output.lower() or "bei-jing" in str_output.lower()
+
+        del model