Files
composable_kernel/example
Amir Ghamarian 86f7ebcf27 Add bs32 narrow tier: 2-warp 16x16 MFMA kBlockM=32 for decode
New unified_attention_decode_bs32_kernel_traits: 2 warps, 16x16 MFMA,
kBlockM=32, kBlockQ=4 for GQA-8 with block_size=32. Uses DecodePolicy
(NumWarpPerGroup=2) for deterministic LDS access.

Replaces the small tier (kBlockQ=8, 87.5% waste) for bs32 tiny decode
with 2x less query waste (kBlockQ=4, 75% waste).

Made-with: Cursor
2026-03-30 17:04:38 +00:00
..
2026-01-14 07:31:45 -08:00