feat(gemm_a8w8_blockscale_bpreshuffle): cold-cache retune triton routing for DeepSeek MLA decode-M

[valarLip]

Summary

Cold-cache re-tune of the a8w8 blockscale B-preshuffle GEMM triton routing for
DeepSeek MLA decode-M shapes on gfx950 (MI355X), plus a test harness that can
actually measure triton, and a units fix in the tuned CSVs.

The routing decisions in #4202 were made with an in-process timing harness that
wrapped the call in a no-arg lambda, which defeats perftest's arg-rotation
cache-busting → every iteration hit a hot L2 → optimistic numbers and, worse,
cold-cache-suboptimal configs. This PR re-measures everything cache-rotated
(tensors passed positionally so perftest rotates the weights) and validates every
change end-to-end through the op-test dispatch.

Changes

1. op_tests/test_gemm_a8w8_blockscale.py — measure triton

• Add a run_triton path + triton us/TFLOPS/TB/s/err and triton/ck columns so
  the harness times triton alongside ck (dispatch) and asm. Previously triton was
  never exercised unless a CSV row already routed to it.
• Remove the -m choices=[...] allowlist so arbitrary decode M (e.g. 48, 80) can be
  benchmarked; the default M list is unchanged.
• Triton column fails loud (no try/except swallow) like the ck/asm columns; when
  ck_preshuffle=False the triton columns are simply omitted (pandas NaN-fills).

2. wqkv_a (N=2048, K=7168) — fix #4202 regression

The #4202 triton configs for wqkv_a were hot-cache tuned (BLOCK_SIZE_M=4/8,
NUM_KSPLIT=4) and are slower than asm cold-cache (13–16µs vs ~10–13µs) — i.e.
#4202 was a regression on this shape as-shipped. Re-tuned cold-cache
(BLOCK_SIZE_M=16, NUM_KSPLIT=8, num_stages=3) → triton now wins 9–16% at
M≤32, and routing extended to M∈{1,2,4,8,16,32,48,64,80,128} (all err=0).

3. wo_b (N=7168) — new triton routing

wo_b was not in #4202. RowParallelLinear(o_groups·o_lora_rank=16384, dim=7168):

• tp1 (K=16384): route M≤64 → triton (M≤32 win 11–15%, M=64 tie, err=0).
• tp4 (K=4096): route M=1..32 → triton (win 9–28%); new tuned JSON added.
• tp8 (K=2048): unchanged — triton loses (shallow K, ck/asm already optimal).

Pattern: triton's split-K advantage scales with K depth (tp1 > tp4 > tp8).

4. Tuned CSVs — unify bw column to TB/s

The bw column mixed units: asm/ck/cktile rows in GB/s (~600–4400), triton rows in
TB/s (~1–5). Converted all asm/ck/cktile rows to TB/s so the column is internally
consistent. us (which drives routing/dedup) and tflops were already consistent.

Not changed (verified correct cold-cache)

• wq_b tp1/4/8 (N=65536/16384/8192, K=1536): feat(gemm_a8w8_blockscale_bpreshuffle): route decode-M DeepSeek shapes to triton #4202 routing is correct
  cold-cache — its shipped configs perform fine; no change needed.

Test plan

• Every routed shape re-run through op_tests/test_gemm_a8w8_blockscale.py
--ck_preshuffle 1; dispatch (ck column) now runs triton for routed M with
  err=0 and matches the standalone triton column.
• Config lookup (get_CKGEMM_config) confirms libtype per M for every shape.
• CSV merge still loads (no duplicate (gfx,cu_num,M,N,K) keys).
• CI: Standard + Triton on MI35X.

[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:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
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 4264 --add-label <label>