[Bugfix][ROCm] Stop running shared experts on the aux stream

[stefankoncarevic]

Purpose

LM Eval Qwen3-5 Models on MI355 fails on Qwen3.5-35B-A3B-DEP2:

FAILED evals/gsm8k/test_gsm8k_correctness.py::test_gsm8k_correctness[Qwen3.5-35B-A3B-DEP2]
AssertionError: GSM8K metric too low: 0.0182 < 0.8400 - 0.0300 = 0.8100

Bisect over the window in which the job turned red lands on 47ececb,
"[Perf][ROCm] Dual-stream decode with hipgraphs" (#48223).

That commit widened the shared-expert multi-stream overlap from
current_platform.is_cuda() to is_cuda_alike(), gated on dp_size > 1, so
the shared experts run on the aux stream on ROCm for the first time. The output
is corrupted there: the server comes up clean and then answers with unparsable
runs of tokens. dp_size > 1 is also why only this config broke — the other two
Qwen3.5 models in the same job run TP2 and never take the path.

This is a workaround, not a fix. It turns the overlap off on ROCm, restoring
the behaviour every ROCm deployment had before #48223, and leaves CUDA
untouched. The new path itself still needs to be repaired, and I am happy to
help with that; what is known so far:

• It needs concurrency. A single greedy request answers correctly.
• It is not tied to capturing the whole decode step: cudagraph_mode=PIECEWISE
  still fails.
• It cannot be observed from Python, since graph replay bypasses that path, and
  instrumentation serialises enough to hide it.

Test Plan

gfx950 (MI355X), 2 GPUs, the failing job as CI runs it, unmodified config and
full question set:

pytest -s -v evals/gsm8k/test_gsm8k_correctness.py \
  --config-list-file=configs/models-qwen35-mi355.txt

Run on current main and again with this change.

Test Result

Before, on current main: Qwen3.5-35B-A3B-DEP2 fails, accuracy 0.0000 with an
invalid rate of 0.705.

After, same config and the full 1319 questions:

GSM8K Results for Qwen/Qwen3.5-35B-A3B:
  Measured metric: 0.8613
  Expected metric: 0.8400
  Tolerance: 0.0300
  Questions: 1319
  Invalid rate: 0.000
✅ GSM8K test passed for Qwen/Qwen3.5-35B-A3B

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[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.

[simondanielsson]

FYI: #52024

[AndreasKaratzas]

we are going to close this one in favor of the revert.