support models from www.modelscope.cn (#994)

Co-authored-by: mulin.lyh <mulin.lyh@taobao.com>

python/sglang/srt/server.py

@@ -74,6 +74,8 @@ from sglang.srt.utils import (
     enable_show_time_cost,
     kill_child_process,
     maybe_set_triton_cache_manager,
+    prepare_model,
+    prepare_tokenizer,
     set_ulimit,
 )
 from sglang.utils import get_exception_traceback
@@ -250,6 +252,10 @@ def launch_server(
     )
     logger.info(f"{server_args=}")
 
+    # Use model from www.modelscope.cn, first download the model.
+    server_args.model_path = prepare_model(server_args.model_path)
+    server_args.tokenizer_path = prepare_tokenizer(server_args.tokenizer_path)
+
     # Launch processes for multi-node tensor parallelism
     if server_args.nnodes > 1:
         if server_args.node_rank != 0:

python/sglang/srt/utils.py

@@ -701,3 +701,23 @@ def add_api_key_middleware(app, api_key):
         if request.headers.get("Authorization") != "Bearer " + api_key:
             return JSONResponse(content={"error": "Unauthorized"}, status_code=401)
         return await call_next(request)
+
+
+def prepare_model(model_path):
+    if "SGLANG_USE_MODELSCOPE" in os.environ:
+        if not os.path.exists(model_path):
+            from modelscope import snapshot_download
+
+            return snapshot_download(model_path)
+    return model_path
+
+
+def prepare_tokenizer(tokenizer_path):
+    if "SGLANG_USE_MODELSCOPE" in os.environ:
+        if not os.path.exists(tokenizer_path):
+            from modelscope import snapshot_download
+
+            return snapshot_download(
+                tokenizer_path, ignore_patterns=["*.bin", "*.safetensors"]
+            )
+    return tokenizer_path