sgl-project/sglang · error · ValueError

Can not get local ip

Error message

Can not get local ip

What it means

get_local_ip_auto tries several strategies (configured IP, NIC lookup, UDP-socket trick to a remote resolver) to determine the host IP. If all fail and no fallback value was supplied, it raises ValueError('Can not get local ip') — the machine effectively has no usable outbound network path.

Source

Thrown at python/sglang/srt/utils/network.py:374

        2. Network interface enumeration via get_local_ip_by_nic()
        3. Remote connection method via get_local_ip_by_remote()
    """
    # Try environment variable
    host_ip = os.getenv("SGLANG_HOST_IP", "") or os.getenv("HOST_IP", "")
    if host_ip:
        return host_ip
    logger.debug("get_ip failed")
    # Fallback
    if ip := get_local_ip_by_nic():
        return ip
    logger.debug("get_local_ip_by_nic failed")
    # Fallback
    if ip := get_local_ip_by_remote():
        return ip
    logger.debug("get_local_ip_by_remote failed")
    if fallback:
        return fallback
    raise ValueError("Can not get local ip")


def get_zmq_socket(
    context: zmq.Context,
    socket_type: zmq.SocketType,
    endpoint: Optional[str] = None,
    bind: bool = True,
) -> Union[zmq.Socket, Tuple[int, zmq.Socket]]:
    """Create and configure a ZeroMQ socket.

    Args:
        context: ZeroMQ context to create the socket from.
        socket_type: Type of ZeroMQ socket to create.
        endpoint: Optional endpoint to bind/connect to. If None, binds to a random TCP port.
        bind: Whether to bind (True) or connect (False) to the endpoint. Ignored if endpoint is None.

    Returns:
        If endpoint is None: Tuple of (port, socket) where port is the randomly assigned TCP port.

View on GitHub (pinned to 0132848349)

Solutions

  1. Set SGLANG_LOCAL_IP (or SGLANG_LOCAL_IP_NIC with netifaces installed) to the bind IP explicitly.
  2. Fix the container/host networking so a default route to the bootstrap address exists.
  3. If calling get_local_ip_auto programmatically, pass fallback="127.0.0.1" for single-node runs.

Example fix

# before
ip = get_local_ip_auto()  # raises on isolated host

# after
import os
os.environ["SGLANG_LOCAL_IP"] = "10.0.0.5"
ip = get_local_ip_auto()
Defensive patterns

Strategy: fallback

Validate before calling

import socket

def has_route() -> bool:
    try:
        s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
        s.connect(("8.8.8.8", 80)); s.close(); return True
    except OSError:
        return False

Try / catch

try:
    ip = get_local_ip_auto()
except ValueError:
    ip = os.environ.get("SGLANG_LOCAL_IP", "127.0.0.1")

Prevention

When it happens

Trigger: Running on an isolated/air-gapped host or a container with no default route, so connecting a UDP socket to the bootstrap/remote address fails, and no SGLANG_LOCAL_IP/NIC env or fallback argument is provided.

Common situations: Docker/K8s pods with restricted networking, CI runners without external access, misconfigured hosts where DNS/route is down. Hits engine init, encoder bootstrap registration, and mooncake transfer-engine init.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/af8d9fd1997724ca. Report an issue: GitHub.