[Kernel] Enable OAI Triton MXFP4 MoE on SM12x (#58877)

Signed-off-by: lazymio <mio@lazym.io>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
This commit is contained in:
lazymio
2026-10-06 01:49:13 -04:00
committed by GitHub
co-authored by Claude Fable 5.1 Shengqi Chen
parent b3082d3044
commit 40c86ae1ce
4 changed files with 96 additions and 7 deletions
@@ -120,7 +120,14 @@ def _swizzle_mxfp4(quant_tensor, scale, num_warps=8):
K = scale.shape[-1]
pad_k = -K % 4
scale = torch.nn.functional.pad(scale, (0, pad_k))
elif current_platform.is_device_capability_family(100):
elif current_platform.is_device_capability_family(
100
) or current_platform.is_device_capability_family(120):
# Native MXFP (cap >= 10.0) requires the persistent kernel. On
# SM12x, epilogue_subtile=1 also keeps the heuristic away from a
# triton_kernels 3.5.1 codegen bug (SWAP_XW with a subtiled
# epilogue); block_m is pinned per call in
# gpt_oss_triton_kernels_moe to fit the 99KB shared memory.
constraints = {
"is_persistent": True,
"epilogue_subtile": 1,