[eble-amd]

Purpose

Merge 9b2f34c from upstream to gfx11 in stages determined by conflicts. My individual commits have brief notes on conflict resolution which are gathered here (and lightly edited) for convenience.

1. from dc2e8f1 ([ROCm][Perf] W4A16: keep skinny GEMM zero-points packed 4-bit vllm-project/vllm#54965)
   Since this work originated here, we're leaving the local branch as it is and just recording a merge.
2. from ceaa9f9
   Conflict resolution: Before, gfx11 had "dflare" in the list of supported methods. Now, it is absent from the list of unsupported methods.
3. from 8c1557a
   Ruff formatting configuration changed upstream.
   • for conflicting hunks, kept gfx11 side and then ran pre-commit
   • hand-edited a benchmark script to copy a function parameter description from one function to related functions
4. from d7e755c
   Choose the upstream versioning changes over gfx11's 8ef3e3c. Also revert the portion of the earlier 6742a6b which 8ef3e3c says didn't suffice.

Validation (agentic)

Base edd1baf (gfx11); PR 0e01145 (upstream through 9b2f34c). Radeon 8060S (gfx1151), ROCm 10.2.0a20261015, torch 2.13.0, Triton 3.8.0. Online serving, batch size 1; base and PR runs alternated in one session. Medians over runs.

Case	Runs base/PR	TTFT ms base	TTFT ms PR	Δ	TPOT ms base	TPOT ms PR	Δ	Acceptance length base	Acceptance length PR
Qwen3.6-35B-A3B W4A16, 4k prefix + 1k text + image	4/3	502.5	503.5	+0.2%	12.30	12.23	-0.5%	—	—
Qwen3.6-35B-A3B W4A16 MTP, MMStar	5/4	199.9	198.1	-0.9%	9.96	10.01	+0.5%	1.849	1.846
Qwen3-30B-A3B AWQ EAGLE3, AIME	2/2	220.2	221.1	+0.4%	13.98	13.98	0.0%	2.594	2.594
Gemma-4-31B-IT AWQ MTP, MMStar	2/2	1228.0	1233.0	+0.4%	36.06	36.10	+0.1%	2.932	2.932
Gemma-4-E4B-IT AWQ DFlare, RoboVQA	2/2	307.3	307.6	+0.1%	6.00	6.01	+0.2%	6.909	6.909
Qwen3-VL-4B-Instruct AWQ, image	2/2	675.7	676.0	0.0%	14.19	14.15	-0.3%	—	—

Server-reported input and output token counts are identical between base and PR in every case.

No regression was found in the tested cases.

Excluded: one run of the PR version in which both Qwen3.6 cases had TTFT 10-14% higher (every request, not just the tail) with unchanged TPOT. The other cases in that run, and all other PR runs of both Qwen3.6 cases, match base.

---

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.

[eble-amd]

There's a dependency issue in the build. I don't know why I didn't hit that in my own builds, but I'll mark this PR as a draft while I investigate.

[eble-amd]

There's a dependency issue in the build. I don't know why I didn't hit that in my own builds, but I'll mark this PR as a draft while I investigate.

I reproduced this when I tried to build just now. This didn't happen yesterday. I'm continuing to investigate.

[eble-amd]

There's a dependency issue in the build. I don't know why I didn't hit that in my own builds, but I'll mark this PR as a draft while I investigate.

I reproduced this when I tried to build just now. This didn't happen yesterday. I'm continuing to investigate.

#1382 fixed that.

Our gfx11 branch has advanced since I first tested this merge. I have merged those changes into this PR. The conflicts arose from changes in comment formatting, so I propose to try to save time by not repeating A/B testing with a new baseline. It's important to press forward with updating gfx11 to catch up with upstream main.

[eble-amd]

The kernel correctness tests have been waiting more than an hour for a runner. I have run them in my local workspace.

Command:

python -m pytest -v tests/kernels/moe/test_hybrid_w4a16_moe.py tests/kernels/quantization/test_hybrid_w4a16_triton.py tests/kernels/quantization/test_rdna_hybrid_w4a16.py tests/kernels/quantization/test_rocm_compressed_tensors_w4a16.py tests/kernels/quantization/test_rocm_skinny_gemms.py tests/kernels/quantization/test_dynamic_int8_lm_head.py tests/kernels/test_wvsplitk_fused_silu.py tests/models/multimodal/processing/test_openvla.py

Output (tail):

======================================= warnings summary =======================================
.venv/lib/python3.12/site-packages/torch/jit/_script.py:365: 14 warnings
  /scratch/eble/vllm/.venv/lib/python3.12/site-packages/torch/jit/_script.py:365: DeprecationWarning: `torch.jit.script_method` is deprecated. Please switch to `torch.compile` or `torch.export`.
    warnings.warn(

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
================== 562 passed, 2492 skipped, 14 warnings in 289.88s (0:04:49) ==================