Use kN0=64 to save vgprs

This commit is contained in:
Qianfeng Zhang
2025-04-22 14:45:27 +00:00
parent 65ddb1a863
commit 26db7e0b7c

View File

@@ -35,7 +35,7 @@ struct HstuAttentionFwdBlockTile<64>
template <>
struct HstuAttentionFwdBlockTile<128>
{
using type = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
using type = ck_tile::sequence<128, 64, 32, 128, 32, 128>;
using gemm0_warps = ck_tile::sequence<4, 1, 1>;
using gemm1_warps = ck_tile::sequence<4, 1, 1>;
};