Skip to content

Support different dtype for LinearCombination #556

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 5 commits into
base: main
Choose a base branch
from
Draft

Support different dtype for LinearCombination #556

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
ef94850
Support different dtype for LinearCombination
joyalbin Oct 13, 2025
64e777a
Correct usage of LinearCombination with proper inputs.
joyalbin Oct 17, 2025
528c1c5
Fix merge conflict
joyalbin Oct 17, 2025
cfac845
Fix benchmarks MixedPrecision testcases
joyalbin Oct 22, 2025
66322cc
Fix merge conflicts
joyalbin Oct 22, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions examples/02_bmg_gemm_mixed_dtype/02_bmg_gemm_f16_u4_f16.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -263,7 +263,7 @@ struct ExampleRunner {
using GEMMDispatchPolicy = cutlass::gemm::MainloopIntelXeXMX16<PipelineStages>;
using EpilogueDispatchPolicy = cutlass::epilogue::IntelXeXMX16;

using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementAccumulator, ElementCompute,
using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementOutput, ElementCompute,
ElementAccumulator, ElementAccumulator, cutlass::FloatRoundStyle::round_to_nearest>;

using FusionCallBacks = cutlass::epilogue::fusion::FusionCallbacks<EpilogueDispatchPolicy, EpilogueOp, TileShape,
Expand Down Expand Up @@ -648,7 +648,7 @@ int main(int argc, const char** argv) {
using GEMMDispatchPolicy = cutlass::gemm::MainloopIntelXeXMX16MixedPrecision<PipelineStages>;
using EpilogueDispatchPolicy = cutlass::epilogue::IntelXeXMX16;

using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementAccumulator, ElementComputeEpilogue,
using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementOutput, ElementComputeEpilogue,
ElementAccumulator, ElementAccumulator, cutlass::FloatRoundStyle::round_to_nearest>;

using FusionCallBacks = cutlass::epilogue::fusion::FusionCallbacks<EpilogueDispatchPolicy, EpilogueOp, TileShape,
Expand Down
4 changes: 2 additions & 2 deletions examples/02_bmg_gemm_mixed_dtype/02_bmg_gemm_f16_u4_s8.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,7 +258,7 @@ struct ExampleRunner {
using GEMMDispatchPolicy = cutlass::gemm::MainloopIntelXeXMX16<PipelineStages>;
using EpilogueDispatchPolicy = cutlass::epilogue::IntelXeXMX16;

using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementAccumulator, ElementCompute,
using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementOutput, ElementCompute,
ElementAccumulator, ElementAccumulator, cutlass::FloatRoundStyle::round_to_nearest>;

using FusionCallBacks = cutlass::epilogue::fusion::FusionCallbacks<EpilogueDispatchPolicy, EpilogueOp, TileShape,
Expand Down Expand Up @@ -735,7 +735,7 @@ int main(int argc, const char** argv) {
using GEMMDispatchPolicy = cutlass::gemm::MainloopIntelXeXMX16MixedPrecision<PipelineStages>;
using EpilogueDispatchPolicy = cutlass::epilogue::IntelXeXMX16;

using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementAccumulator, ElementComputeEpilogue,
using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementOutput, ElementComputeEpilogue,
ElementAccumulator, ElementAccumulator, cutlass::FloatRoundStyle::round_to_nearest>;

using FusionCallBacks = cutlass::epilogue::fusion::FusionCallbacks<EpilogueDispatchPolicy, EpilogueOp, TileShape,
Expand Down
2 changes: 1 addition & 1 deletion examples/10_bmg_grouped_gemm_mixed_dtype/10_bmg_grouped_gemm_f16_u4.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,7 +115,7 @@ int main(int argc, const char** argv) {
using GEMMDispatchPolicy = cutlass::gemm::MainloopIntelXeXMX16GroupMixedPrecision<PipelineStages>;
using EpilogueDispatchPolicy = cutlass::epilogue::IntelXeXMX16Group;

using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementAccumulator, ElementComputeEpilogue,
using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementOutput, ElementComputeEpilogue,
ElementAccumulator, ElementAccumulator, cutlass::FloatRoundStyle::round_to_nearest>;

using FusionCallBacks = cutlass::epilogue::fusion::FusionCallbacks<EpilogueDispatchPolicy, EpilogueOp, TileShape,
Expand Down
2 changes: 1 addition & 1 deletion examples/10_bmg_grouped_gemm_mixed_dtype/bmg_grouped_gemm_mixed_dtype_runner.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -389,7 +389,7 @@ struct ExampleRunner {
using GEMMDispatchPolicy = cutlass::gemm::MainloopIntelXeXMX16<PipelineStages>;
using EpilogueDispatchPolicy = cutlass::epilogue::IntelXeXMX16;

using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementAccumulator, ElementCompute,
using EpilogueOp = cutlass::epilogue::fusion::LinearCombination<ElementOutput, ElementCompute,
ElementAccumulator, ElementAccumulator, cutlass::FloatRoundStyle::round_to_nearest>;

using FusionCallBacks = cutlass::epilogue::fusion::FusionCallbacks<EpilogueDispatchPolicy, EpilogueOp, TileShape,
Expand Down
6 changes: 3 additions & 3 deletions include/cutlass/epilogue/collective/xe_array_epilogue.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,7 +115,7 @@ class CollectiveEpilogue<
static constexpr FloatRoundStyle RoundStyle = FloatRoundStyle::round_to_nearest;

static_assert(cute::is_same_v<typename FusionCallbacks::Operation,
fusion::LinearCombination<ElementAccumulator, ElementCompute, ElementSource, ElementScalar, RoundStyle>>,
fusion::LinearCombination<ElementOutput, ElementCompute, ElementSource, ElementScalar, RoundStyle>>,
"Only Linear Combination Epilogue is supported for Grouped GEMM at the moment.");

static constexpr int SubgroupSize = DispatchPolicy::SubgroupSize;
Expand Down Expand Up @@ -411,7 +411,7 @@ class CollectiveEpilogue<
cst_callbacks.begin();

auto acc_frag = recast<Array<ElementCompute, FragmentSize>>(accumulators);
auto trD_compute_frag = recast<Array<ElementCompute, FragmentSize>>(trD_compute);
auto trD_compute_frag = recast<Array<ElementOutput, FragmentSize>>(trD_compute);

Tensor trD = make_tensor<ElementOutput>(Shape<Int<FragmentSize>>{});
auto trD_frag = recast<Array<ElementOutput, FragmentSize>>(trD);
Expand Down Expand Up @@ -445,7 +445,7 @@ class CollectiveEpilogue<
if constexpr (is_destination_supported) {
CUTLASS_PRAGMA_UNROLL
for (int i = 0; i < size(trD_compute_frag); ++i) {
trD_frag(i) = cutlass::NumericArrayConverter<ElementOutput, ElementCompute, FragmentSize>{}(trD_compute_frag(i));
trD_frag(i) = cutlass::NumericArrayConverter<ElementOutput, ElementOutput, FragmentSize>{}(trD_compute_frag(i));
}
copy(params.xe_store_d.with(get<1>(load_store_tensors)), trD, tCgD(_, epi_m, epi_n));
}
Expand Down
4 changes: 2 additions & 2 deletions include/cutlass/epilogue/collective/xe_epilogue.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -389,7 +389,7 @@ class CollectiveEpilogue<
cst_callbacks.begin();

auto acc_frag = recast<Array<ElementCompute, FragmentSize>>(accumulators);
auto trD_compute_frag = recast<Array<ElementCompute, FragmentSize>>(trD_compute);
auto trD_compute_frag = recast<Array<ElementOutput, FragmentSize>>(trD_compute);

Tensor trD = make_tensor<ElementOutput>(Shape<Int<FragmentSize>>{});
auto trD_frag = recast<Array<ElementOutput, FragmentSize>>(trD);
Expand Down Expand Up @@ -423,7 +423,7 @@ class CollectiveEpilogue<
if constexpr (is_destination_supported) {
CUTLASS_PRAGMA_UNROLL
for (int i = 0; i < size(trD_compute_frag); ++i) {
trD_frag(i) = cutlass::NumericArrayConverter<ElementOutput, ElementCompute, FragmentSize>{}(trD_compute_frag(i));
trD_frag(i) = cutlass::NumericArrayConverter<ElementOutput, ElementOutput, FragmentSize>{}(trD_compute_frag(i));
}
copy(params.xe_store_d, trD, tCgD(_, epi_m, epi_n));
}
Expand Down
4 changes: 2 additions & 2 deletions test/unit/gemm/device/default_gemm_configuration.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1465,7 +1465,7 @@ struct DefaultGemmConfigurationToCutlass3Types<
cutlass::gemm::collective::KernelScheduleAuto
>::CollectiveOp;

using EpilogueOp = epilogue::fusion::LinearCombination<float, float>;
using EpilogueOp = epilogue::fusion::LinearCombination<ElementOutput, float>;

using FusionCallBacks = cutlass::epilogue::fusion::FusionCallbacks<
epilogue::IntelXeXMX16,
Expand Down Expand Up @@ -1567,7 +1567,7 @@ struct DefaultGemmConfigurationToCutlass3Types<
cutlass::gemm::collective::KernelScheduleAuto
>::CollectiveOp;

using EpilogueOp = epilogue::fusion::LinearCombination<float, float>;
using EpilogueOp = epilogue::fusion::LinearCombination<ElementOutput, float>;

using FusionCallBacks = cutlass::epilogue::fusion::FusionCallbacks<
epilogue::IntelXeXMX16,
Expand Down
2 changes: 1 addition & 1 deletion test/unit/gemm/device/default_gemm_group_configuration.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,7 +87,7 @@ struct DefaultGemmGroupConfiguration<

using TiledMma = typename CollectiveMainloop::TiledMma;

using EpilogueOp = epilogue::fusion::LinearCombination<float, float>;
using EpilogueOp = epilogue::fusion::LinearCombination<ElementOutput, float>;

using FusionCallBacks = epilogue::fusion::FusionCallbacks<
epilogue::IntelXeXMX16Group,
Expand Down
2 changes: 1 addition & 1 deletion test/unit/gemm/device/xe_gemm_fp16_fp16_fp16_tensor_op_fp32.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,7 +51,7 @@ struct XE_Device_Gemm_fp16_fp16_f16_tensor_op_f32 {
cute::half_t, LayoutA,
cute::half_t, LayoutB,
float, layout::RowMajor,
cute::half_t>;
float>;

using Gemm = gemm::device::GemmUniversalAdapter<gemm::kernel::GemmUniversal<
cute::Shape<int,int,int,int>,
Expand Down