[ROCm][Bugfix] Restrict multi-stream shared experts to gfx950

[djramic]

Purpose

#56098 enabled the MoE shared-experts auxiliary stream across ROCm. On gfx942, that causes Elastic EP scaling to hang: AMD CI build 12870 failed three eager parameterizations of test_elastic_ep_scaling, and enforce_eager_light fails 5 of 5 runs locally on main.

Tracing the rescale through StatelessGroupCoordinator and PyNcclCommunicator shows that all four ranks of the new DP group return from ncclCommInitRank, but the first all_reduce never completes and c10d times out after 300 seconds. With
the aux stream disabled, the same collective completes within 35 ms; ncclCommInitRank takes about 10 seconds in both cases.

This change restricts shared-expert auxiliary-stream overlap on ROCm to gfx950, so #56098's behavior is preserved there. The temporary gate prevents the gfx942 Elastic EP scaling failure until the interaction between the aux stream and RCCL initialization is resolved.

Test Plan

Four MI300X GPUs:

cd /vllm-workspace/tests
pytest -v -s distributed/test_elastic_ep.py

Test Result

Before: 0 of 5 runs of test_elastic_ep_scaling[enforce_eager_light] passed,
each failing after a 300 s [c10d] waitForInput timeout followed by
Scale failed: The actor died unexpectedly before finishing this task.

After: the full file passes, 5 of 5 tests in 26m30s.

AI assistance was used.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.