[Misc][ModelScope] Change to use runtime VLLM_USE_MODELSCOPE (#18655)

Signed-off-by: Mengqing Cao <cmq0113@163.com>
Signed-off-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Isotr0py <2037008807@qq.com>

tests/test_regression.py

@@ -60,6 +60,9 @@ def test_model_from_modelscope(monkeypatch: pytest.MonkeyPatch):
     # model: https://modelscope.cn/models/qwen/Qwen1.5-0.5B-Chat/summary
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_MODELSCOPE", "True")
+        # Don't use HF_TOKEN for ModelScope repos, otherwise it will fail
+        # with 400 Client Error: Bad Request.
+        m.setenv("HF_TOKEN", "")
         llm = LLM(model="qwen/Qwen1.5-0.5B-Chat")
 
         prompts = [

vllm/model_executor/model_loader/default_loader.py

@@ -11,8 +11,8 @@ import torch
 from torch import nn
 from transformers.utils import SAFE_WEIGHTS_INDEX_NAME
 
+from vllm import envs
 from vllm.config import LoadConfig, LoadFormat, ModelConfig, VllmConfig
-from vllm.envs import VLLM_USE_MODELSCOPE
 from vllm.logger import init_logger
 from vllm.model_executor.model_loader.base_loader import BaseModelLoader
 from vllm.model_executor.model_loader.utils import (
@@ -64,7 +64,7 @@ class DefaultModelLoader(BaseModelLoader):
 
         Returns the path to the downloaded model, or None if the model is not
         downloaded from ModelScope."""
-        if VLLM_USE_MODELSCOPE:
+        if envs.VLLM_USE_MODELSCOPE:
             # download model from ModelScope hub,
             # lazy import so that modelscope is not required for normal use.
             # pylint: disable=C.

vllm/transformers_utils/__init__.py

 # SPDX-License-Identifier: Apache-2.0
 
-from vllm.envs import VLLM_USE_MODELSCOPE
+from vllm import envs
 
-if VLLM_USE_MODELSCOPE:
+if envs.VLLM_USE_MODELSCOPE:
     try:
         # Patch here, before each import happens
         import modelscope

vllm/transformers_utils/config.py

@@ -24,7 +24,7 @@ from transformers.models.auto.modeling_auto import (
     MODEL_FOR_CAUSAL_LM_MAPPING_NAMES)
 from transformers.utils import CONFIG_NAME as HF_CONFIG_NAME
 
-from vllm.envs import VLLM_USE_MODELSCOPE
+from vllm import envs
 from vllm.logger import init_logger
 # yapf conflicts with isort for this block
 # yapf: disable
@@ -45,13 +45,12 @@ from vllm.transformers_utils.configs import (ChatGLMConfig, Cohere2Config,
 from vllm.transformers_utils.utils import check_gguf_file
 from vllm.utils import resolve_obj_by_qualname
 
-if VLLM_USE_MODELSCOPE:
+if envs.VLLM_USE_MODELSCOPE:
     from modelscope import AutoConfig
 else:
     from transformers import AutoConfig
 
 MISTRAL_CONFIG_NAME = "params.json"
-HF_TOKEN = os.getenv('HF_TOKEN', None)
 
 logger = init_logger(__name__)
 
@@ -130,7 +129,7 @@ def list_repo_files(
             ]
         # if model is remote, use hf_hub api to list files
         try:
-            if VLLM_USE_MODELSCOPE:
+            if envs.VLLM_USE_MODELSCOPE:
                 from vllm.transformers_utils.utils import (
                     modelscope_list_repo_files)
                 return modelscope_list_repo_files(repo_id,
@@ -185,7 +184,7 @@ def file_or_path_exists(model: Union[str, Path], config_name: str,
     return file_exists(str(model),
                        config_name,
                        revision=revision,
-                       token=HF_TOKEN)
+                       token=os.getenv('HF_TOKEN', None))
 
 
 def patch_rope_scaling(config: PretrainedConfig) -> None:
@@ -312,7 +311,7 @@ def get_config(
             model,
             revision=revision,
             code_revision=code_revision,
-            token=HF_TOKEN,
+            token=os.getenv('HF_TOKEN', None),
             **kwargs,
         )
 
@@ -324,7 +323,7 @@ def get_config(
                 model,
                 revision=revision,
                 code_revision=code_revision,
-                token=HF_TOKEN,
+                token=os.getenv('HF_TOKEN', None),
                 **kwargs,
             )
         else:
@@ -334,7 +333,7 @@ def get_config(
                     trust_remote_code=trust_remote_code,
                     revision=revision,
                     code_revision=code_revision,
-                    token=HF_TOKEN,
+                    token=os.getenv('HF_TOKEN', None),
                     **kwargs,
                 )
             except ValueError as e:
@@ -352,7 +351,7 @@ def get_config(
                     raise e
 
     elif config_format == ConfigFormat.MISTRAL:
-        config = load_params_config(model, revision, token=HF_TOKEN, **kwargs)
+        config = load_params_config(model, revision, **kwargs)
     else:
         supported_formats = [
             fmt.value for fmt in ConfigFormat if fmt != ConfigFormat.AUTO
@@ -561,7 +560,7 @@ def get_sentence_transformer_tokenizer_config(model: str,
             # If model is on HuggingfaceHub, get the repo files
             repo_files = list_repo_files(model,
                                          revision=revision,
-                                         token=HF_TOKEN)
+                                         token=os.getenv('HF_TOKEN', None))
         except Exception:
             repo_files = []
 
@@ -768,7 +767,7 @@ def get_hf_image_processor_config(
     **kwargs,
 ) -> dict[str, Any]:
     # ModelScope does not provide an interface for image_processor
-    if VLLM_USE_MODELSCOPE:
+    if envs.VLLM_USE_MODELSCOPE:
         return dict()
     # Separate model folder from file path for GGUF models
     if check_gguf_file(model):

vllm/transformers_utils/tokenizer.py

@@ -13,7 +13,7 @@ import huggingface_hub
 from transformers import (AutoTokenizer, PreTrainedTokenizer,
                           PreTrainedTokenizerFast)
 
-from vllm.envs import VLLM_USE_MODELSCOPE
+from vllm import envs
 from vllm.logger import init_logger
 from vllm.lora.request import LoRARequest
 from vllm.transformers_utils.tokenizer_base import (TokenizerBase,
@@ -168,7 +168,7 @@ def get_tokenizer(
 ) -> AnyTokenizer:
     """Gets a tokenizer for the given model name via HuggingFace or ModelScope.
     """
-    if VLLM_USE_MODELSCOPE:
+    if envs.VLLM_USE_MODELSCOPE:
         # download model from ModelScope hub,
         # lazy import so that modelscope is not required for normal use.
         # pylint: disable=C.