perf(gptoss): tune bf16 GEMM shapes to non-hipblaslt kernels

[valarLip]

Problem

Starting gpt-oss-120b with -tp 2 --enable-dp-attention --enable-expert-parallel --enable-tbo crashes during CUDA/HIP graph capture:

Hip error: 'operation not permitted when stream is capturing'(900) at hipblaslt.cpp:147
Engine Core: load model runner failed

Root cause

Several gpt-oss bf16 GEMM shapes have no tuned config, so they fall back to the torch path (F.linear → hipBLASLt). Under --enable-tbo, decode batches are split into bs/2 micro-batches, producing GEMM shapes (e.g. m=2880 n=128 k=4096, the o/down projection at the TBO ubatch token count) that are executed for the first time during cudagraph capture. hipBLASLt's algorithm/workspace init on first touch of a shape is stream-capture-illegal → HIP error 900.

Confirmed via HIPBLASLT_LOG_MASK=32 trace: the crashing matmul carries a hipBLASLt solution_index, i.e. it went through the hipBLASLt path.

Fix

Tune the affected gpt-oss bf16 GEMM shapes (all dtype=bf16 outdtype=bf16) so they resolve to direct-launch kernels instead of torch/hipBLASLt:

N	K	role
2880	4096	o / down proj
5120	2880	qkv / gate-up
128	2880	router gate
201088	2880	lm_head

Result: all resolve to asm / flydsl / triton / opus — no torch/hipblaslt. Entries live in the per-model model_configs/gptoss_bf16_tuned_gemm.csv (not the shared bf16_tuned_gemm.csv), sorted by gfx with gfx1250 last.

Tuned with gemm_a16w16_tune.py --libtype asm,opus,flydsl,triton,skinny (hipblaslt and torch excluded).

Verification (gfx950, MI355X)

• Server starts and passes cudagraph capture — no HIP error 900.
• GSM8K (chat, num_concurrent=65, num_fewshot=3): flexible-extract 0.872–0.884 across runs (threshold 0.87, baseline 0.9) — accuracy preserved.

[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 4352 --add-label <label>