Bind the MessageQueue remote socket atomically

[metamergebot]

Motivation

MessageQueue.__init__ sets up the remote XPUB socket in two steps: get_open_port() probes for a free port, then remote_socket.bind() claims it. Nothing holds the port in between. When several writers with remote readers are created concurrently on one host (for example, multi-node model-parallel groups that each build a broadcaster during initialization), two of them can be handed the same port, and the later one dies in __init__:

  File ".../sglang/srt/distributed/device_communicators/shm_broadcast.py", line 244, in __init__
    self.remote_socket.bind(address)
zmq.error.ZMQError: Address already in use (addr='tcp://[::1]:50003')

The failing rank takes its whole process group down with it. #29455 closed this window for the local socket; the remote socket still has it. #15903 proposed moving both sockets onto get_zmq_socket as a wider refactor, but it has been in conflict with main for months.

Modifications

• shm_broadcast.py: replace the probe-then-bind pair for the remote socket with pyzmq's bind_to_random_port, which binds atomically and retries on EADDRINUSE. The window starts at the port get_open_port() returned and spans 32 ports, so SGLANG_PORT still anchors the range and the ephemeral fallback is unchanged. Exhausting the window raises ZMQBindError. IPV6 is still set before binding, and the exported handle carries the port that was actually bound.
• The local-socket path is untouched. No new flag.

Accuracy Tests

Not applicable: only the way the remote broadcast socket picks its port at initialization changes.

Speed Tests and Profiling

Not applicable: initialization only, nothing on the request path.

Test plan

CPU-only script outside the repo that imports the real shm_broadcast and network modules, run against this branch and against main, for IPv4 and IPv6:

• Race made deterministic: get_open_port is patched to return a port another socket already holds. main raises ZMQError(EADDRINUSE) from __init__; this branch binds another port inside the window.
• Four writer processes started concurrently with the same starting port: on main three of the four fail with EADDRINUSE; on this branch all four bind distinct ports inside the window.
• Writer plus remote reader round trip through the bound socket (wait_until_ready, enqueue, dequeue) still delivers.
• Every port in the window occupied raises ZMQBindError.
• ruff, ruff-format, isort and codespell at the versions pinned in .pre-commit-config.yaml pass on the changed file.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests. (No new test file, matching shm_broadcast: retry bind on EADDRINUSE (fix dp-attention port race) #29455; the deterministic checks above are described in the test plan.)
• Update documentation according to Write documentations. (Not applicable.)
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed. (Not applicable.)
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): 🚫 Run #37401987293
Latest PR Test (Extra): ❌ Run #37402003034
Latest PR Test (AMD ROCm 10): ❌ Run #37401987257

[metamergebot]

This port clears our bar on dc56fa4b7f084476e7ec0576ee60733826da2842: every base-a and base-b job is green.
base-c and the accelerator suites are reported but not gated, because they run
on scarce hardware and upstream merges through them routinely.

gate 20✓✗2✂4 lint✓ build 3✓✗1✂3 A✓ B✓ C 0/8⏳ other 6✓✂6 · accel ✗amd,npu,xpu

@merrymercy @charlotte12l — could one of you take a look and merge it if you are happy? Nothing else is waiting on our side.