[jiacao-amd]

Purpose

On gfx950, wvSplitK picks YTILE 3 or 4 for wide outputs. Once N >= 5 those tiles are up to 4x slower than YTILE 2, so N=5 fell behind hipBLASLt on shapes like an LM head shard (e.g. the draft/target LM head under EAGLE with K=3 and 5 rows).

This PR:

• uses YTILE 2 for N >= 5 on gfx950 (N <= 4 and other GPUs are unchanged);
• lets gfx950 run N=6 through wvSplitK where it beats hipBLASLt (M > 1024, 2048 <= K <= 8192).

Kernel results (MI355X, bf16, µs per call, CUDA graph over rotating weights)

LM head shard M=50016, K=6144:

N	hipBLASLt	wvSplitK before	wvSplitK after
4	129.9	103.0	103.0
5	129.4	385.8	95.1
6	123.7	— (hipBLASLt)	98.1

Over a 630-shape grid (M 16–100032, K 512–16384, N 1–6):

• N=5: 35% faster than before on average (geomean), 42/105 shapes more than 20% faster, none slower beyond noise.
• N=6 in the enabled range: 13% faster than hipBLASLt on average (geomean), worst case +7% at M=100032 where both are bandwidth bound.
• N <= 4: unchanged.
• N=7/8 beat hipBLASLt on only part of the grid and stay on hipBLASLt.

Max relative error vs an fp32 reference: 0.4%.

E2E (MI355X, MiniMax-M3-MXFP4, TP4, EAGLE3 K=3)

vLLM nightly 8e2beaa3 (AITER v0.1.26) plus #61889, which that nightly needs to start MiniMax-M3; the baseline is the same image without this PR. AgenticX: InferenceX recipe minimaxm3/vllm/mi355x-fp4-mtp/agentic.yaml, 3600 s per point, one run each (run-to-run noise is about ±3%). Cells are baseline → this PR.

AgenticX (selected points; also measured: the previous revision at TP2 C1–C16)

Conc	Throughput/GPU (tok/s)	P50 interactivity	P90 interactivity	Mean TPOT	P90 TTFT
20	33,664 → 33,891 (+0.7%)	201.3 → 205.0 (+1.8%)	129.8 → 128.4 (−1.1%)	5.57 → 5.43 ms (−2.5%)	1.14 → 1.06 s (−6.4%)
24	40,256 → 40,332 (+0.2%)	169.9 → 170.7 (+0.4%)	97.5 → 98.1 (+0.6%)	6.80 → 6.72 ms (−1.2%)	1.19 → 1.14 s (−4.0%)

The change applies when the draft LM head runs with 5 or 6 rows, about a quarter of the time at C20–C24 and less at higher concurrency, which matches the 1–2.5% lower TPOT.

Accuracy (GSM8K), InferenceX gsm8k.yaml, 5-shot, all 1,319 questions, real EAGLE3 acceptance, thinking on:

	strict-match	flexible-extract
baseline	96.21%	96.13%
this PR	95.91%	95.83%

Standard error ±0.53%.

Test plan

• tests/model_executor/layers/test_rocm_unquantized_gemm.py: new dispatch cases for N=5/6/7 on gfx950 and other gfx9 (42 passed, 2 skipped on MI355X).
• Kernel timing and accuracy over the grid above, old vs new wvSplitK and hipBLASLt.

🤖 Generated with Claude Code

[jiacao-amd]

Closing: on MiniMax-M3.1-Flash (the target now) decode batches never hit the N=5/6 wvSplitK path, and the related lm_head dispatch change measured only ~0.6% end to end.