[roberteg16]

Summary

Conflict-free upstream batch: the 15 commits between 49fb2ee348 and d785eb51ce (2026-08-18 01:33 UTC .. 2026-08-18 21:46 UTC).

67 files, +1856/−245. No deleted files.

Boundary probed in a ladder: all coarse offsets conflict (k=25 .. k=447); stepping from zero, k=1..15 are clean and k=16 conflicts.

The next commit, 5fa8ca971a ([ROCm][CI] Move ROCm AITER quantization tests,
vllm-project#40938), conflicts and gets its own PR.

Merge commit only — do not squash or rebase.

AI assistance was used to prepare this merge.

Test plan

• pre-commit run --all-files green on the merged tree
• Build + 6-model benchmark sweep vs the Strix Halo dashboard — no regression on transformers 5.15.0, all six within the gate (results)

[roberteg16]

Six-model verification sweep — no regression

Run against the CI wheel built from this branch, with transformers 5.15.0
(the version this stack unpins in batch 11 / #1258).

item	value
wheel	vllm-0.26.1rc1.dev2126+gf3e3a2a45 (run 34202804814, branch rogarcia.merge-upstream-121)
transformers	5.15.0
torch / torchvision / torchaudio	2.13.0 / 0.28.0 / 2.11.0, all +rocm10.1.0a20260906
triton	3.8.0+git4cff872c.rocm10.1.0a20260906
GPU	AMD Radeon 8060S (gfx1151)

Baselines are the Strix Halo reference run regression_20260907_093058 — the
same board, not the worked example in the skill docs.

bench	test id	decode tok/s	baseline	Δ
bench1	Qwen3-30B-A3B-Instruct-2507-AWQ-4bit_128	80.8	80.2	+0.7%
bench2	Qwen3.6-35B-A3B-W4A16-llmcompressor_VLM_dyn-lm-int8g32	83.2	82.9	+0.4%
bench3	Gemma-3-4B-IT_VLM_w4a16	61.3	61.6	−0.5%
bench4	Qwen3-Omni-30B-A3B-Instruct_VLM_AWQ-4bit	76.8	75.5	+1.7%
bench5	Qwen2.5-0.5B-Instruct_128	359.9	357.0	+0.8%
bench6	Gemma-4-31B-IT_VLM_AWQ-4bit	10.3	10.3	+0.0%

All six: 10/10 successful requests, 0 failures. Every benchmark is within the 1%
decode gate; four are above baseline.

This covers #1269 (batch 22), #1270 (batch 23), #1271 (batch 24) and #1272
(batch 25): the wheel is built from -121, whose history contains all four.

Kernel tests

Run against the same wheel, from a neutral working directory so the installed
wheel is exercised rather than the source tree:

tests/kernels/core/test_rocm_aiter_ops.py
tests/kernels/quantization/test_triton_scaled_mm.py
-> 61 passed, 12 skipped

The 61 passing include test_scaled_mm_td_matches_plain, which exercises the
use_td keyword whose signature had to be preserved when resolving the batch-2
conflict.

The 12 skips are test_rocm_aiter_ops.py, the tests #1271 adds, and they
cannot run on this board. require_aiter() gates on
is_aiter_found_and_supported(), which is get_cdna_version() > 2 — CDNA 3 or
newer. gfx1151 is RDNA 3.5, so the predicate returns False by design, not for
want of an install. Confirmed at runtime:

arch: AMD_Radeon_8060S
is_aiter_found_and_supported(): False

Installing aiter from PyPI does not help and is actively misleading: that
package (aiter==0.13.20191203) is an unrelated homonym with no rms_norm or
rmsnorm2d_fwd_with_add. It was removed again. #1271's tests need CDNA 3+
hardware (MI3xx) to be meaningful — they are not verified here, and the
checkbox for them is left unticked.

AI assistance was used to prepare this verification.