ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-29 19:28:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

change the gemm1 tile from 64x128x128 to 128x64x128

Browse Source
This commit is contained in:
mtgu0705
2025-05-27 10:46:43 -05:00
parent 52b764d59f
commit 2f0ee8ccb1
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
example/67_gemm_microscaling/moe_gemm1_xdl_mx_fp4_bns.cpp
View File

@@ -147,8 +147,8 @@ constexpr ck::index_t ScaleBlockSize = 32; // scaling block
constexpr ck::index_t KPerBlock = 256 / DataPackedSize; // 256 f4 = 128 fp4x2
static constexpr ck::index_t Nswizzle = false;
static constexpr ck::index_t ActOP = 0; // 0: gelu_and_mul, 1: silu_and_mul
static constexpr ck::index_t MPerBlock = 64;
static constexpr ck::index_t NPerBlock = 128;
static constexpr ck::index_t MPerBlock = 128;
static constexpr ck::index_t NPerBlock = 64;
static constexpr ck::index_t BlockSize = 256;
static constexpr bool MulRoutedWeight = true;
@@ -161,7 +161,7 @@ using DeviceOpInstance = ck::tensor_operation::device::Devic
MPerBlock, NPerBlock, KPerBlock,
16, 16,
16, 16,
2, 4,
4, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
2, 2, S<1, 32, 1, 8>, S<8, 1, 1, 1>,