ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-19 02:01:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

update MX moe GEMM1 hotloopscheduling

Browse Source
This commit is contained in:
mtgu0705
2025-05-27 08:49:46 -05:00
parent a36a747e29
commit c8a329e271
2 changed files with 12 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
example/67_gemm_microscaling/moe_gemm1_xdl_mx_fp4_bns.cpp
View File

@@ -159,7 +159,7 @@ using DeviceOpInstance = ck::tensor_operation::device::Devic
AElementOp, BElementOp, CDEElementOp, GemmSpec,
ScaleBlockSize, BlockSize,
MPerBlock, NPerBlock, KPerBlock,
16, 16,
16, 16,
16, 16,
2, 4,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
@@ -237,6 +237,12 @@ int main(int argc, char* argv[])
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({sorted_tile_num + 1}));
max_token_id.mData[0] = valid_size;
if(tokens * topk > valid_size)
{
printf("err config, tokens * topk > valid_size\n");
exit(-1);
}
for(int i = 0; i < sorted_tile_num; i++)
{
expert_ids.mData[i] = i / ck::math::integer_divide_ceil(valid_tile_num, experts);