ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-16 08:44:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

change default config

Browse Source
This commit is contained in:
coderfeli
2025-02-20 15:02:56 +00:00
parent 6c09a5970d
commit 4cab3b5fb2
2 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
example/65_gemm_multiply_multiply/moe_gemm1.cpp
View File

@@ -132,8 +132,8 @@ using AElementOp = PassThrough;
using BElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t MPerBlock = 192;
static constexpr ck::index_t MXDLPerWave = 3;
static constexpr ck::index_t MPerBlock = 128;
static constexpr ck::index_t MXDLPerWave = 2;
static constexpr ck::index_t NXDLPerWave = 2;
static constexpr ck::index_t BLOCKSIZE = 256;
static constexpr ck::index_t NPerBlock = 128;

5
example/65_gemm_multiply_multiply/moe_gemm2.cpp
View File

@@ -171,7 +171,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemm
// MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
MXDLPerWave, 1, S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, true, false, A0DataType>;
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, false, false, A0DataType>;
// kernel 2: 128->32x128x128
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, EDataType>;
@@ -241,7 +241,8 @@ int main(int argc, char* argv[])
Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1}));
max_token_id.mData[0] = valid_size;
// max_token_id.mData[0] = valid_size;
max_token_id.mData = {valid_size, 0, 2, 3, 4, 6, 8, 10, 12, 13};
int eids[] = {0, 0,1, 2,3, 3, 4,4, 5, 5, 6, 7, 7, 3, 3, 3};
for (int i = 0; i < sorted_tile_num; i++) {
expert_ids.mData[i] = eids[i];