Bugfix: fix broken of download models from modelscope

tests/test_regression.py

@@ -52,6 +52,25 @@ def test_gc():
     allocated = torch.cuda.memory_allocated()
     assert allocated < 50 * 1024 * 1024
 
+def test_model_from_modelscope(monkeypatch):
+    # model: https://modelscope.cn/models/qwen/Qwen1.5-0.5B-Chat/summary
+    MODELSCOPE_MODEL_NAME = "qwen/Qwen1.5-0.5B-Chat"
+    monkeypatch.setenv("VLLM_USE_MODELSCOPE", "True")
+    try:
+        llm = LLM(model=MODELSCOPE_MODEL_NAME)
+
+        prompts = [
+            "Hello, my name is",
+            "The president of the United States is",
+            "The capital of France is",
+            "The future of AI is",
+        ]
+        sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+        outputs = llm.generate(prompts, sampling_params)
+        assert len(outputs) == 4
+    finally:
+        monkeypatch.delenv("VLLM_USE_MODELSCOPE", raising=False)
 
 if __name__ == "__main__":
     import pytest

vllm/config.py

@@ -112,7 +112,11 @@ def __init__(
         self.revision = revision
         self.code_revision = code_revision
         self.rope_scaling = rope_scaling
-        self.tokenizer_revision = tokenizer_revision
+        # The tokenizer version is consistent with the model version by default.
+        if tokenizer_revision is None:
+            self.tokenizer_revision = revision
+        else:
+            self.tokenizer_revision = tokenizer_revision
         self.quantization = quantization
         self.quantization_param_path = quantization_param_path
         self.enforce_eager = enforce_eager

vllm/transformers_utils/config.py

@@ -1,7 +1,8 @@
 from typing import Dict, Optional
 
-from transformers import AutoConfig, PretrainedConfig
+from transformers import PretrainedConfig
 
+from vllm.envs import VLLM_USE_MODELSCOPE
 from vllm.logger import init_logger
 from vllm.transformers_utils.configs import (ChatGLMConfig, DbrxConfig,
                                              JAISConfig, MPTConfig, RWConfig)
@@ -24,6 +25,10 @@ def get_config(model: str,
                code_revision: Optional[str] = None,
                rope_scaling: Optional[dict] = None) -> PretrainedConfig:
     try:
+        if VLLM_USE_MODELSCOPE:
+            from modelscope import AutoConfig
+        else:
+            from transformers import AutoConfig
         config = AutoConfig.from_pretrained(
             model,
             trust_remote_code=trust_remote_code,