ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-19 22:39:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix and improve the gemm quant pipeline infrastructure (#3245)

Browse Source
This commit is contained in:
Thomas Ning
2025-11-26 18:04:27 -08:00
committed by GitHub
parent 79aae7c7f7
commit a38aeceb21
11 changed files with 96 additions and 272 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
test/ck_tile/gemm_block_scale/test_gemm_quant_typed.cpp
View File

@@ -88,11 +88,7 @@ using BQuantTypes = ::testing::Types<
std::tuple<RowMajor, ColumnMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigBase, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigBase, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigBase, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigBase, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigBase, GroupSize2D128N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigBase, GroupSize2D128N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigBase, GroupSize2D128N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigBase, GroupSize2D128N>
std::tuple<RowMajor, ColumnMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigBase, GroupSize2D64N>
>;
// clang-format on

7
test/ck_tile/grouped_gemm_multi_d/test_grouped_gemm_multi_d_util.hpp
View File

@@ -124,7 +124,12 @@ class TestCkTileGroupedGemmMultiD : public ::testing::Test
using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV3<GemmPipelineProblem>;
using BaseGemmPipeline = std::conditional_t<
Config::Pipeline_ == (PipelineType::Memory),
ck_tile::BaseGemmPipelineAgBgCrMem<GemmPipelineProblem>,
std::conditional_t<Config::Pipeline_ == (PipelineType::CompV3),
ck_tile::BaseGemmPipelineAgBgCrCompV3<GemmPipelineProblem>,
ck_tile::BaseGemmPipelineAgBgCrCompV4<GemmPipelineProblem>>>;
const ck_tile::index_t k_grain = gemm_descs[0].k_batch * Config::K_Tile_;
const ck_tile::index_t K_split =