fix(tuned_gemm): gate skinny GEMM default to archs with a real kernel

[yichiche]

Summary

• is_skinny_default_shape() could select the skinny GEMM path on architectures where those kernels are compiled as assert(false) stubs, aborting at runtime.
• Restrict the skinny default to the archs that actually have a kernel body: gfx90a / gfx942 / gfx950.
• Unsupported archs (e.g. gfx1250) now fall through to the regular GEMM dispatch instead of aborting.

Motivation

get_GEMM_A16W16_config() falls back to a default config when a shape is not in bf16_tuned_gemm.csv. One of those fallbacks is the skinny path:

elif is_skinny_default_shape(M, N, K, dtype, cu_num):
    default_config["libtype"] = "skinny"

The skinny GEMM kernels — wvSpltK, LLMM1, wv_splitk_small_fp16_bf16 — only have a real body inside #if defined(__HIP__MI350_MI300_MI250__) in csrc/kernels/custom_kernels.cu. That macro is defined only for __gfx90a__ || __gfx942__ || __gfx950__ (custom_kernels.cu:28-30); every other arch compiles the #else branch, whose body is UNREACHABLE_CODE, i.e. assert(false);.

is_skinny_default_shape() itself only looked at M/N/K/dtype/cu_num, never at the arch. So on an arch without the kernel body, any untuned skinny-shaped GEMM (small M, N <= cu_num, K % 8 == 0) selects a stub and aborts. This is reachable on gfx1250 with a bf16 model whose small-M shapes are not yet in the tuned config.

Modifications

aiter/tuned_gemm.py: add an arch allowlist and check it first in is_skinny_default_shape().

_SKINNY_GEMM_ARCHS = {"gfx90a", "gfx942", "gfx950"}

def is_skinny_default_shape(M, N, K, dtype, cu_num=None):
    if get_gfx() not in _SKINNY_GEMM_ARCHS:
        return False
    ...

The set mirrors the __HIP__MI350_MI300_MI250__ definition exactly, and follows the same pattern already used a few lines above for _OPUS_WS_ARCHS (aiter/tuned_gemm.py:77).

Impact

• gfx90a / gfx942 / gfx950: no behaviour change — get_gfx() returns one of the allowlisted values, so the predicate result is identical to before.
• Other archs: the skinny default is no longer selected; those shapes fall through to the existing non-skinny dispatch. This replaces a runtime abort with a working kernel.
• Tuned configs are unaffected — this only touches the untuned-shape default path.

Test

• Functional fix, no perf claim: the change removes an abort on archs that never had a working skinny kernel, and is a no-op on the three archs that do.
• Verified get_gfx() returns a bare arch string from GFX_MAP (aiter/jit/utils/chip_info.py:66-68), so the membership test matches the existing get_gfx() == "gfx950" style comparisons elsewhere in the repo.
• ruff / black / isort clean.

[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
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 4791 --add-label <label>