[charlotte12l]

Summary

When SGLANG_PORT is pinned and multiple MessageQueue writers start
concurrently on the same node, the XPUB local socket setup uses
get_open_port() followed by a later bind() — a non-atomic
check-then-bind. Two writers can be handed the same port and then collide
on bind() with EADDRINUSE.

One scheduler subprocess dies at init:

zmq.error.ZMQError: Address already in use (addr='tcp://127.0.0.1:50000')
  File ".../sglang/srt/distributed/parallel_state.py", line 464, in __init__
    self.mq_broadcaster = MessageQueue.create_from_process_group(self.cpu_group, 1 << 22, 6)
  File ".../sglang/srt/distributed/device_communicators/shm_broadcast.py", line 214, in __init__
    self.local_socket.bind(socket_addr)

which then cascades into a gloo failure on every rank:

RuntimeError: [gloo/transport/tcp/pair.cc] Connection reset by peer
  File ".../sglang/srt/model_executor/model_runner.py", line 1189, in get_available_gpu_memory
    torch.distributed.all_reduce(tensor, op=ReduceOp.MIN, group=group)

This replaces get_open_port() + bind() with zmq's bind_to_random_port,
which picks and binds a free port atomically. When SGLANG_PORT is set the
search starts at that base (min=base, max=base + 8) to preserve the
existing port range; otherwise it falls back to the default ephemeral range.

A similar atomic fix is proposed upstream in sgl-project#15885 (PR sgl-project#15903).

Test

Concurrent MessageQueue writers under a pinned SGLANG_PORT no longer
hit the EADDRINUSE error; each binds a distinct port.

---

CI States

Latest PR Test (Base): ❌ Run #28311918658
Latest PR Test (Extra): ❌ Run #28311918565

[charlotte12l]

Retargeting to upstream sgl-project/sglang.