Upstream sync 7/N: merge cf567cbc71 attention benchmarks (conflict)

[roberteg16]

Context

Seventh step of the batched upstream catch-up. Stacked on #1098 (base rogarcia.merge-upstream-06); retarget to gfx11 once #1097 and #1098 land.

Conflict-only step: exactly one upstream commit.

Merged upstream commit	cf567cbc71 "[Attention] Improve attention benchmarks: configs and profiling (vllm-project#39336)"
Landed on upstream main	2026-06-12 20:24 UTC
Commits in this PR	1

It lands squarely in benchmarks/attention_benchmarks/, where the fork carries six commits of its own (skip/intermittent/cooldown flags, gfx1151 TRITON_ATTN tuning, YAML configs for pytest).

Conflicts and resolutions

common.py — upstream adds a median_time field to BenchmarkResult.to_dict(); the fork nulls every timing field when the run was skipped. Kept the fork's if not self.skip else None guard and applied it to median_time as well, so skipped runs stay uniformly null.

benchmark.py — both sides add imports at the same line (json vs os/shutil/subprocess). Kept both.

A second breakage git did not flag

create_skipped_result() is a fork-only helper, and upstream made median_time a required dataclass field with no default. The helper does not pass it, so every skipped benchmark would have raised TypeError — and git merged that file region without a murmur, because the two sides touched different parts of it.

Fixed here, and all seven BenchmarkResult(...) call sites in the directory were audited with an AST pass against the dataclass's required-field set; this was the only one missing an argument.

This is the second occurrence of the same pattern in this sync series (see #1097's stale ops.impl("ggml_*") registrations). Worth carrying forward as a standing check: after each merge, verify that fork-only code still satisfies any API upstream changed.

How bad this would have been. Not a rarely-used manual flag: tests/kernels/attention/benchmark/test_benchmark_attention.py defines a per-platform SKIP_CASES dict, passes it to benchmark.py via --skip, and the golden validator relies on the skipped case still being present in the output JSON (if actual.get("skip"): continue). So the attention golden-regression suite hits this path on every run that has skip cases defined for the platform — it would have failed outright on gfx1151, not just for someone experimenting with --skip by hand.

Test plan

• ruff check and ruff format --check clean on the resolved files. (The README.md formatting complaint reproduces on upstream's own copy of the file and is not covered by the repo's ruff-format hook.)
• AST audit of every BenchmarkResult(...) construction against the required-field set.
• ROCm build + the four gfx1151 benchmarks — run on the stacked Upstream sync 8/N: merge a3195fab7b (57 commits, conflict-free) #1101, which contains this merge. All pass, no regression: decode flat, TTFT within ±1.4%. Full table on Upstream sync 8/N: merge a3195fab7b (57 commits, conflict-free) #1101.

[mgehre-amd]

Why do we have create_skipped_result on the fork? Do we still need it?

[roberteg16]

Why do we have create_skipped_result on the fork? Do we still need it?

We need it as long as long we kept our feature of golden reference testing