[Misc] Normalize --host parameter to use plain hostname without scheme (#19309)

Co-authored-by: 墨楼 <huangzhilin.hzl@antgroup.com>
Co-authored-by: Liangsheng Yin <lsyincs@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>

python/sglang/test/few_shot_gsm8k.py

@@ -14,7 +14,12 @@ import numpy as np
 
 from sglang.lang.api import set_default_backend
 from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
-from sglang.utils import download_and_cache_file, dump_state_text, read_jsonl
+from sglang.utils import (
+    download_and_cache_file,
+    dump_state_text,
+    normalize_base_url,
+    read_jsonl,
+)
 
 INVALID = -9999999
 
@@ -46,7 +51,7 @@ def get_answer_value(answer_str):
 
 def run_eval(args):
     # Select backend
-    set_default_backend(RuntimeEndpoint(f"{args.host}:{args.port}"))
+    set_default_backend(RuntimeEndpoint(normalize_base_url(args.host, args.port)))
 
     if args.data_path is None:
         # Read data
@@ -142,7 +147,7 @@ if __name__ == "__main__":
     parser.add_argument("--num-questions", type=int, default=200)
     parser.add_argument("--max-new-tokens", type=int, default=512)
     parser.add_argument("--parallel", type=int, default=128)
-    parser.add_argument("--host", type=str, default="http://127.0.0.1")
+    parser.add_argument("--host", type=str, default="127.0.0.1")
     parser.add_argument("--port", type=int, default=30000)
     parser.add_argument("--temperature", type=float, default=0.0)
     args = parser.parse_args()

python/sglang/test/test_utils.py

@@ -44,7 +44,7 @@ from sglang.srt.utils import (
     retry,
 )
 from sglang.test.run_eval import run_eval
-from sglang.utils import get_exception_traceback
+from sglang.utils import get_exception_traceback, normalize_base_url
 
 # General test models
 DEFAULT_MODEL_NAME_FOR_TEST = "meta-llama/Llama-3.1-8B-Instruct"
@@ -377,7 +377,7 @@ def call_select_guidance(context, choices, model=None):
 
 def add_common_other_args_and_parse(parser: argparse.ArgumentParser):
     parser.add_argument("--parallel", type=int, default=64)
-    parser.add_argument("--host", type=str, default="http://127.0.0.1")
+    parser.add_argument("--host", type=str, default="127.0.0.1")
     parser.add_argument("--port", type=int, default=None)
     parser.add_argument(
         "--backend",
@@ -426,7 +426,7 @@ def auto_config_device() -> str:
 
 def add_common_sglang_args_and_parse(parser: argparse.ArgumentParser):
     parser.add_argument("--parallel", type=int, default=64)
-    parser.add_argument("--host", type=str, default="http://127.0.0.1")
+    parser.add_argument("--host", type=str, default="127.0.0.1")
     parser.add_argument("--port", type=int, default=30000)
     parser.add_argument("--backend", type=str, default="srt")
     parser.add_argument(
@@ -450,7 +450,7 @@ def select_sglang_backend(args: argparse.Namespace):
     if args.backend.startswith("srt"):
         if args.backend == "srt-no-parallel":
             global_config.enable_parallel_encoding = False
-        backend = RuntimeEndpoint(f"{args.host}:{args.port}")
+        backend = RuntimeEndpoint(normalize_base_url(args.host, args.port))
     elif args.backend.startswith("gpt-"):
         backend = OpenAI(args.backend)
     else:
@@ -459,14 +459,15 @@ def select_sglang_backend(args: argparse.Namespace):
 
 
 def _get_call_generate(args: argparse.Namespace):
+    base_url = normalize_base_url(args.host, args.port)
     if args.backend == "lightllm":
-        return partial(call_generate_lightllm, url=f"{args.host}:{args.port}/generate")
+        return partial(call_generate_lightllm, url=f"{base_url}/generate")
     elif args.backend == "vllm":
-        return partial(call_generate_vllm, url=f"{args.host}:{args.port}/generate")
+        return partial(call_generate_vllm, url=f"{base_url}/generate")
     elif args.backend == "srt-raw":
-        return partial(call_generate_srt_raw, url=f"{args.host}:{args.port}/generate")
+        return partial(call_generate_srt_raw, url=f"{base_url}/generate")
     elif args.backend == "outlines":
-        return partial(call_generate_outlines, url=f"{args.host}:{args.port}/generate")
+        return partial(call_generate_outlines, url=f"{base_url}/generate")
     elif args.backend == "guidance":
         from guidance import models
 
@@ -479,10 +480,11 @@ def _get_call_generate(args: argparse.Namespace):
 
 
 def _get_call_select(args: argparse.Namespace):
+    base_url = normalize_base_url(args.host, args.port)
     if args.backend == "lightllm":
-        return partial(call_select_lightllm, url=f"{args.host}:{args.port}/generate")
+        return partial(call_select_lightllm, url=f"{base_url}/generate")
     elif args.backend == "vllm":
-        return partial(call_select_vllm, url=f"{args.host}:{args.port}/generate")
+        return partial(call_select_vllm, url=f"{base_url}/generate")
     elif args.backend == "guidance":
         from guidance import models
 

python/sglang/utils.py

@@ -12,6 +12,7 @@ import sys
 import time
 import traceback
 import urllib.request
+import warnings
 import weakref
 from collections import OrderedDict
 from concurrent.futures import ThreadPoolExecutor
@@ -122,6 +123,19 @@ def dump_state_text(filename: str, states: list, mode: str = "w"):
             )
 
 
+def normalize_base_url(host: str, port: int) -> str:
+    if host.startswith("http://") or host.startswith("https://"):
+        warnings.warn(
+            f"Including the scheme in --host ('{host}') is deprecated. "
+            f"Pass just the hostname (e.g. '127.0.0.1') instead.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+    else:
+        host = f"http://{host}"
+    return f"{host}:{port}"
+
+
 class HttpResponse:
     def __init__(self, resp):
         self.resp = resp