sgl-project/sglang · critical · RuntimeError

Server process exited with code {return_code}

Error message

Server process exited with code {return_code}

What it means

wait_for_http_ready's _raise_if_process_exited polled the server process and found it already terminated with a known exit code, so readiness waiting aborts instead of hanging on a dead server.

Source

Thrown at python/sglang/utils.py:542

    Terminate the process and automatically release the reserved port.
    """
    from sglang.srt.utils import kill_process_tree

    kill_process_tree(process.pid)

    lock_socket = process_socket_map.pop(process, None)
    if lock_socket is not None:
        release_port(lock_socket)


def _raise_if_process_exited(process: Optional[Any]) -> None:
    if process is None:
        return

    if hasattr(process, "poll"):
        return_code = process.poll()
        if return_code is not None:
            raise RuntimeError(f"Server process exited with code {return_code}")
        return

    if hasattr(process, "is_alive") and not process.is_alive():
        return_code = getattr(process, "exitcode", None)
        if return_code is None:
            raise RuntimeError("Server process exited")
        raise RuntimeError(f"Server process exited with code {return_code}")


def _is_wait_timeout(start_time: float, timeout: Optional[int]) -> bool:
    if timeout is None:
        return False
    return time.perf_counter() - start_time > timeout


def wait_for_http_ready(
    url: str,
    timeout: Optional[int] = None,

View on GitHub (pinned to 0132848349)

Solutions

1. Reproduce the server launch manually and read its stdout/stderr for the root cause
2. Fix the underlying startup failure (model path, GPU memory, flags)
3. Retain and print the process output in the test harness for diagnosis

try:
    wait_for_http_ready(server_process, url)
except RuntimeError as e:
    if 'exited with code' in str(e):
        print(server_process.stdout.read())  # surface root cause

When it happens

Trigger: Server subprocess crashes during startup (bad args, OOM, CUDA error, import failure) while the test harness polls /health.

Common situations: Benchmarks/tests calling launch_server_cmd then wait_for_http_ready; server exits due to invalid model path or GPU error.

Related errors

• Server process exited
• Initialization failed. Please see the error messages above.
• Failed to get server info. {error_data['error']['message']}
• world_size ({world_size}) is less than tensor_parallel_degre
• Could not connect to remote scheduler at {self.server_args.s

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/010295ce8fe0c8ca. Report an issue: GitHub.