[TRITON] gfx1201: gemm_a8w8 tuning configs (Mistral-3 / Qwen3 shapes)

[carlushuang]

Summary

Two small additions that let aiter run on RDNA4 (gfx1201, RX 9070 XT family) without the calling project having to maintain its own kernel/config replicas.

1. aiter.ops.triton.activation.silu_and_mul — a triton implementation of the existing HIP silu_and_mul, with the same (out, x) signature so callers can dispatch by arch without changing call sites. The HIP kernel does not compile on RDNA4: its inner activation_kernels.cu uses v_pk_mul_f32, an instruction that exists only on CDNA (gfx9*) and gfx1250.

2. 5 gfx1201-GEMM-A8W8*.json tuning configs for the per-tensor FP8 gemm_a8w8 triton kernel. Without these, gemm_config_utils falls through to the cross-arch default (GROUP_SIZE_M=4), which leaves 75% of M-dim launch slots idle on RDNA4 at decode bs=1..32. Each config is hand-tuned on RX 9070 XT for one of the four projection shapes used by Mistral-3-8B / Qwen3-8B-FP8 (qkv, o, gate_up, down).

Headline numbers (gfx1201, RX 9070 XT, ROCm 7.x)

silu_and_mul triton vs torch fallback (the only other option on gfx1201):

Shape (M, 2H)	Triton	Torch	Δ
(8, 28672)	9.1us	10.0us	-9%
(32, 28672)	9.2us	13.2us	-30%
(1024, 28672)	153us	205us	-25%

gemm_a8w8 per-shape configs vs the cross-arch default at decode bs=1:

Shape	Default	Tuned
qkv	163us	33us
o	45us	28us
gate_up	229us	211us
down	107us	36us

Test plan

• silu_and_mul: bf16 + fp16, 2H non-power-of-2 included; relative err <1% vs F.silu(a)*b (triton accumulates in fp32, so it is in fact more accurate than the bf16 reference)
• gemm_a8w8: get_gemm_config(\"GEMM-A8W8\", M, N, K) returns the new specialized blocks for all 4 (N, K); kernel output 0 abs-err vs BF16 reference (dequant FP8 then matmul) at bs in {1, 8, 32}
• No regression risk for other archs — both additions are arch-specific files / new symbols; nothing existing is renamed or removed

Context

Used by ROCm/ATOM in ROCm/ATOM#749 (gfx1201 / Mistral-3-8B + Qwen3-8B-FP8 enablement). Once this PR lands and the aiter pin in ATOM is bumped, ATOM can delete its _silu_mul_triton and _gfx1201_gemm_a8w8_config replicas.

[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 3168 --add-label <label>

[Chi-Chu319]

Hi, can this be a duplicate of #2592?

[brunomazzottiamd]

Hi @carlushuang. Can you please edit the PR description, removing the content related to SiLU + Mul Triton kernel? Thank you.