[eble-amd]

Purpose

Merge 238cb2b from upstream to gfx11 in stages dictated by conflicts.

• from dc89fdf
  Resolved a conflict in vllm/profiler/wrapper.py by taking the upstream change and manually reapplying gfx11's change to set max_name_column_width=100.
• from 186a1e6
  In vllm/v1/attention/backends/flash_attn.py, upstream changed a block that is indented in gfx11.
• from d43bb2f
  Resolve an insert-insert conflict in tests/test_config.py by keeping both.
• from 9dcf6bf
  Resolved a conflict in the DFlash speculator by keeping the local logic that avoids rebuilding the attention metadata in the full-replay case, and applying the upstream change to the else branch.

Test Plan & Result (agentic)

Base 4685ba1 (gfx11); PR 8f3d9cb (upstream through 238cb2b). Radeon 8060S (gfx1151), ROCm 10.2.0a20261014, torch 2.13.0, Triton 3.8.0.

Cases were chosen where the merged changes touch the decode path on gfx1151: speculative decoding in Model Runner V2 and the Triton top-k/top-p sampler. All cases use greedy decoding at batch size 1.

Qwen3.6-35B-A3B-W4A16 MTP, MMStar

Mean of 4 runs per side, alternating base and PR.

Metric	Base	PR	Δ
Mean TPOT (ms)	10.19	10.13	-0.6%
Median TPOT (ms)	10.14	10.18	+0.3%
Median TTFT (ms)	199.3	199.9	+0.3%
Output throughput (tok/s)	78.47	78.72	+0.3%
Acceptance length	1.851	1.854	+0.2%

Mean TPOT of a single run was either 10.07-10.10 ms or 10.28-10.29 ms on both sides.

Gemma-4-31B-IT AWQ-4bit MTP, MMStar

Mean of 4 runs per side, alternating base and PR.

Metric	Base	PR	Δ
Median TPOT (ms)	36.31	36.30	-0.0%
Median TTFT (ms)	1240.3	1245.2	+0.4%
Output throughput (tok/s)	17.58	17.56	-0.1%
Acceptance length	2.896	2.896	0.0%

apply_top_k_top_p_triton

Wall time per call at batch size 1. GPU time is unchanged for every case with top-k; top-p alone is limited by kernel launch overhead.

Vocab	Filter	Base (µs)	PR (µs)	Δ
151936	top-k 20	476.4	480.3	+0.8%
151936	top-p 0.95	80.1	98.9	+23%
151936	top-k 20, top-p 0.95	571.6	579.5	+1.4%
262144	top-k 20	778.9	773.9	-0.6%
262144	top-p 0.95	91.3	102.3	+12%
262144	top-k 64, top-p 0.95	997.7	996.3	-0.1%

Summary

No regression was found in the tested cases. Top-p-only sampling costs up to 19 µs more per call at batch size 1, which is 0.2% of a 10 ms decode step.

Not tested: Qwen3.6 DFlare (V1 model runner), which fails on both base and PR because of a known module-level aiter import in rocm_aiter_fa.py (fixed upstream in vllm-project#54805).

---

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.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[mgehre-amd]

Thanks!

[eble-amd]

Important: This merge defined a new CI job that is enabled in forks, and it is failing for this PR.

I've sought advice from upstream maintainers: vllm-project#56365 (comment)

• I would prefer not to delay my work of bringing gfx11 up to date.
• This check is optional: it doesn't prevent merging our PRs.
• It is possible that the failure signals something useful to us, say, don't trust a release in cases X, Y, and Z if this check failed. I don't want to disable it until I understand more about it.
• I suspect that the failure will appear and disappear from time to time as we progress through the upstream history.

Therefore, I propose merging this with the new job still enabled, in spite of the failure, and waiting a little while for advice from the upstream maintainers before finally deciding how to handle it in this branch.