Upstream sync 10/N: merge 457a5f34db (31 commits, conflict-free)

[roberteg16]

Summary

Conflict-free upstream batch: the 31 commits between dc5101fb1b and 457a5f34db (2026-08-11 04:43 UTC .. 2026-08-11 16:00 UTC). git merge reported no conflicts; nothing in this merge is a manual resolution.

111 files, +11076/−1080.

Findings

Boundary probed in a ladder: k=25 clean; k=100 through the tip (k=720) all conflict. Stepping forward from 25, k=26..31 are clean and k=32 conflicts.

Deleted files. One, vllm/models/inkling/nvidia/ops/fa4_warmup.py, removed by upstream's 6c95a641e9 ("[2/N][Feat][Perf] Add new warmup infrastructure for JITs…", vllm-project#49315) which migrates Inkling FA4 to the new warmup infrastructure. Verified this is the NVIDIA copy only and that nothing imports it: no fa4_warmup references remain anywhere under vllm/models/inkling/nvidia/, and the directory now contains fa4_rel_attention.py, lamport.py, mm_towers.py, norm.py, qkvr_prep.py, sconv.py and __init__.py. The AMD copy the fork depends on, vllm/models/inkling/amd/ops/fa4_warmup.py, is untouched and its importer vllm/models/inkling/amd/attention.py:42 still resolves.

The next commit, c65aa2ee6c (Bump Transformers version to 5.15.0, vllm-project#51668), conflicts in two requirements files 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
• Deleted NVIDIA fa4_warmup confirmed unreferenced; AMD copy verified intact
• Build + 6-model benchmark sweep vs the Strix Halo dashboard — run 3512149, no regression (see comment)

[roberteg16]

Benchmarks

Six-model sweep run on the Strix Halo board via Custom Regression run 3512149.

No regression — decode and TTFT both flat across the set.

bench	model	decode tok/s	baseline	Δ	TTFT ms	baseline	Δ
bench1	Qwen3-30B-A3B-Instruct-2507-AWQ-4bit	80.2	80.2	+0.1%	187.3	187.0	+0.2%
bench2	Qwen3.6-35B-A3B-W4A16 (int8:g32 lm-head)	82.9	83.0	−0.1%	265.8	281.5	−5.6%
bench3	gemma-3-4b-it-quantized.w4a16	61.6	61.5	+0.2%	464.2	463.8	+0.1%
bench4	Qwen3-Omni-30B-A3B-Instruct-AWQ-4bit	75.5	75.9	−0.6%	688.3	702.1	−2.0%
bench5	Qwen2.5-0.5B-Instruct-AWQ	357.0	358.5	−0.4%	16.9	16.8	+0.2%
bench6	gemma-4-31B-it-AWQ-4bit	10.3	10.3	+0.1%	1536.0	1533.7	+0.2%

Largest decode move is −0.6%, inside the 1% gate. All six report PASS with sanity checks green and exit code 0.

Decode is derived from median_tpot_ms rather than the report's max_output_tokens_per_s, which is rounded to whole tok/s — at bench5's ~357 tok/s one integer step is already 0.3%, so that field cannot resolve a 1% gate.

Baselines are the nightly dashboard run of 2026-09-05 (20260905_180640, vLLM c84454ff4).

Scope. The run built 4f03b854e, GitHub's merge commit for #1259 (9686e281df into 0a7aeafd4d) — the top of batch 12, so it contains this PR and every batch below it, and excludes batch 13 (#1260). Individual batches were not swept separately; a regression introduced by one and offset by another would not be visible here, though nothing in the numbers suggests that.

Environment: host xconucstrhalo07, gfx1151, 96 GiB. ROCm 10.1.0a20260906, torch 2.13.0, triton 3.8.0+git4cff872c, transformers 5.14.1.

AI assistance was used to run this verification.