ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix quant scale matrix layout for block scale gemm (#3079)

Browse Source 
* Adding support for TiledPermuteN

* Adding test

* moving shuffle functions to common place

* resolving commit hook

* fix formatting
This commit is contained in:
Khushbu Agarwal
2025-10-27 13:56:07 -07:00
committed by GitHub
parent a46b725992
commit b11f53a484
10 changed files with 35 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
include/ck_tile/ops/gemm_quant/kernel/gemm_quant_kernel.hpp
View File

@@ -686,8 +686,8 @@ struct QuantGemmKernel
static_assert(std::is_same_v<BQLayout, tensor_layout::gemm::ColumnMajor>);
return make_naive_tensor_view<address_space_enum::global>(
bq_ptr,
make_tuple(kargs.N, kargs.QK_B),
make_tuple(kargs.stride_BQ, 1),
make_tuple(kargs.QK_B, kargs.N),
make_tuple(1, kargs.stride_BQ),
number<GemmPipeline::GetVectorSizeBQ()>{},
number<1>{});
}
@@ -905,9 +905,9 @@ struct QuantGemmKernel
static_assert(std::is_same_v<BQLayout, tensor_layout::gemm::ColumnMajor>);
return make_tile_window(
bq_pad_view,
make_tuple(number<TilePartitioner::NPerBlock>{},
number<TilePartitioner::KPerBlock / GemmPipeline::QuantGroupSize>{}),
{i_n, 0});
make_tuple(number<TilePartitioner::KPerBlock / GemmPipeline::QuantGroupSize>{},
number<TilePartitioner::NPerBlock>{}),
{0, i_n});
}
else
{