CI: use 'sglang serve' in CI tests (#18597)

Co-authored-by: Mick <mickjagger19@icloud.com>
Co-authored-by: sglang-bot <sglangbot@gmail.com>

python/sglang/cli/utils.py

@@ -26,6 +26,33 @@ def _get_lock(model_name_or_path: str, cache_dir: Optional[str] = None):
     return lock
 
 
+def _hf_hub_download(
+    repo_id: str,
+    filename: str,
+    local_dir: Optional[str] = None,
+    **kwargs,
+) -> str:
+    """Unified hf_hub_download supporting both Hugging Face Hub and ModelScope."""
+    if envs.SGLANG_USE_MODELSCOPE.get():
+        from modelscope import model_file_download
+
+        return model_file_download(
+            model_id=repo_id,
+            file_path=filename,
+            cache_dir=local_dir,
+            **kwargs,
+        )
+    else:
+        from huggingface_hub import hf_hub_download as _hf_hub_download_impl
+
+        return _hf_hub_download_impl(
+            repo_id=repo_id,
+            filename=filename,
+            local_dir=local_dir,
+            **kwargs,
+        )
+
+
 # Copied and adapted from hf_diffusers_utils.py
 def _maybe_download_model(
     model_name_or_path: str, local_dir: str | None = None, download: bool = True
@@ -44,8 +71,6 @@ def _maybe_download_model(
         Local directory path that contains the downloaded config file, or the original local directory.
     """
 
-    from sglang.multimodal_gen.runtime.utils.hf_diffusers_utils import hf_hub_download
-
     if os.path.exists(model_name_or_path):
         logger.debug("Model already exists locally")
         return model_name_or_path
@@ -62,7 +87,7 @@ def _maybe_download_model(
                 source_hub,
                 model_name_or_path,
             )
-            file_path = hf_hub_download(
+            file_path = _hf_hub_download(
                 repo_id=model_name_or_path,
                 filename="model_index.json",
                 local_dir=local_dir,
@@ -79,7 +104,7 @@ def _maybe_download_model(
                 source_hub,
                 model_name_or_path,
             )
-            file_path = hf_hub_download(
+            file_path = _hf_hub_download(
                 repo_id=model_name_or_path,
                 filename="config.json",
                 local_dir=local_dir,

python/sglang/multimodal_gen/utils.py

@@ -11,7 +11,6 @@ import inspect
 import math
 import os
 import signal
-import socket
 import sys
 import threading
 import traceback
@@ -23,7 +22,6 @@ from typing import Any, TypeVar, cast
 import cloudpickle
 import torch
 import yaml
-from remote_pdb import RemotePdb
 from torch.distributed.fsdp import MixedPrecisionPolicy
 
 import sglang.multimodal_gen.envs as envs
@@ -546,15 +544,6 @@ class TypeBasedDispatcher:
         raise ValueError(f"Invalid object: {obj}")
 
 
-# For non-torch.distributed debugging
-def remote_breakpoint() -> None:
-    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
-        s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
-        s.bind(("localhost", 0))  # Let the OS pick an ephemeral port.
-        port = s.getsockname()[1]
-        RemotePdb(host="localhost", port=port).set_trace()
-
-
 @dataclass
 class MixedPrecisionState:
     param_dtype: torch.dtype | None = None

python/sglang/test/test_utils.py

@@ -879,18 +879,22 @@ def popen_launch_server(
 
     use_mixed_pd_engine = not pd_separated and num_replicas is not None
     if pd_separated or use_mixed_pd_engine:
-        command = "sglang.launch_pd_server"
+        command = [
+            "python3",
+            "-m",
+            "sglang.launch_pd_server",
+            "--model-path",
+            model,
+            *[str(x) for x in other_args],
+        ]
     else:
-        command = "sglang.launch_server"
-
-    command = [
-        "python3",
-        "-m",
-        command,
-        "--model-path",
-        model,
-        *[str(x) for x in other_args],
-    ]
+        command = [
+            "sglang",
+            "serve",
+            "--model-path",
+            model,
+            *[str(x) for x in other_args],
+        ]
 
     if pd_separated or use_mixed_pd_engine:
         command.extend(["--lb-host", host, "--lb-port", port])
@@ -1308,7 +1312,7 @@ def run_score_benchmark(
                         json=warmup_data,
                         timeout=aiohttp.ClientTimeout(total=30),
                     )
-                except:
+                except Exception:
                     pass  # Ignore warmup errors
 
         test_requests = []
@@ -1376,17 +1380,9 @@ def run_embeddings_benchmark(
 
     async def _run_benchmark():
 
-        # Load tokenizer for generating test data
-        from sglang.srt.utils.hf_transformers_utils import get_tokenizer
-
-        tokenizer = get_tokenizer(model)
-
         def generate_text_with_token_count(num_tokens):
             """Generate text with precise token count using special tokens."""
-            # Use a token that reliably produces 1 token
             special_token = "<|im_start|>"
-            # Verify it's a single token
-            test_tokens = tokenizer.encode(special_token, add_special_tokens=False)
             text = special_token * num_tokens
             return text
 
@@ -1406,7 +1402,7 @@ def run_embeddings_benchmark(
                         json=warmup_data,
                         timeout=aiohttp.ClientTimeout(total=30),
                     )
-                except:
+                except Exception:
                     pass  # Ignore warmup errors
 
         test_requests = []
@@ -1822,7 +1818,7 @@ def run_mulit_request_test(
                     },
                 },
             )
-            ret = response.json()
+            response.json()
 
         with ThreadPoolExecutor(2) as executor:
             list(executor.map(run_one, list(range(4))))