[Config] gptoss bf16 tuned gemm: drop the losing large-M QKV rows on gfx950 (tuned rows lose 16-27% to the hipBLASLt fallback)

[alexnails]

What

The gfx950 rows in gptoss_bf16_tuned_gemm.csv for the gpt-oss QKV projection shape (N=5120, K=2880, bias) at M=4096/8192/16384 pick kernels that lose to the plain hipBLASLt fallback on MI355X:

M (lookup)	tuned pick	hipBLASLt (torch.addmm)	Δ
61366 → 8192 row	flydsl, 1.744 ms	1.271 ms	−27%
4065 → 4096 row	flydsl		−16.5%
16384 row	"triton auto"		−22%

The o-projection shape — which has no tuned row — already runs at 93% of the measured 1424 TF/s GEMM ceiling through the same fallback, which is the tell. Delete the three rows so large-M lookups fall through to hipBLASLt; the small-M decode rows (M ≤ 64) keep flydsl, which wins there (+11.5% at M=64).

Verified on MI355X: the shapes take the hipBLASLt path with outputs identical to torch.addmm (max_abs 0.0), worth ~16 ms/request on a 61k-token gpt-oss prefill (36 launches). Suggests a broader guard for the tuning flow: reject a candidate row at table-generation time if it does not beat the untuned fallback for that shape.

🤖 Generated with Claude Code

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
multigpu	Aiter multi-GPU tests on the 8-GPU runner
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 5159 --add-label <label>

PR title tags & labels:
Component tags ([Triton/Gluon], [HIP], [CK], [ASM], ...) are added to the PR title and as PR labels automatically from the changed files and re-synced on every push — change-type tags like [fix]/[Perf], op tags like [MLA], and human labels (ci:*) are left untouched. Add the no-auto-title label to opt this PR out.

[yifehuan]

LGTM

[github-actions]

No activity for 15 days, so this is now labelled stale. A push or a comment clears it; keep-open exempts it. Nothing is closed automatically.