leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 27 Packages Projects Releases Wiki Activity

v4.1 release

Browse Source
This commit is contained in:
Junkai-Wu
2025-07-03 08:07:53 -04:00
committed by GitHub
parent b995f93317
commit a1aaf2300a
155 changed files with 18407 additions and 6068 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/80_blackwell_geforce_sparse_gemm/80a_blackwell_geforce_mxfp8_bf16_sparse_gemm.cu
+6 -12
View File
@@ -108,7 +108,7 @@ using KernelScheduleType = cutlass::gemm::KernelSparseTmaWarpSpecializedMxf8f6f
using ThreadBlockShape = Shape<_128,_128,_256>; // Threadblock's tile size
using ClusterShape = Shape<_1,_1,_1>; // Shape of the threadblocks in a cluster
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
ArchTag, OperatorClass,
ArchTag, OperatorClass,
ThreadBlockShape, ClusterShape,
cutlass::epilogue::collective::EpilogueTileAuto,
ElementAccumulator, ElementAccumulator,
@@ -175,13 +175,7 @@ cutlass::HostTensor<ElementD, cutlass::layout::PackedVectorLayout> block_referen
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
template <typename T>
auto make_iterator(T* ptr) {
using namespace cute;
if constexpr (cute::is_subbyte_v<T>) {
return subbyte_iterator<T>(ptr);
}
else {
return ptr;
}
return cute::recast_ptr<T>(ptr);
}
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Testbed utility types
@@ -289,7 +283,7 @@ bool initialize_block(
}
cutlass::reference::host::TensorFillRandomUniform(
view, seed, scope_max, scope_min, 0);
return true;
}
/// Initialize blocks that released to sparse Matrix A and its metadata E
@@ -465,7 +459,7 @@ template <typename Gemm>
int run(Options &options)
{
// Initialization
if(!initialize(options))
if(!initialize(options))
{
std::cerr << " Initialization failed! " << std::endl;
exit(-1);
@@ -527,9 +521,9 @@ int main(int argc, char const **args) {
cudaDeviceProp props;
int current_device_id;
CUDA_CHECK(cudaGetDevice(&current_device_id));
CUDA_CHECK(cudaGetDeviceProperties(&props, current_device_id));
if (!(props.major == 12 && props.minor == 0)) {
std::cerr << "This example requires a GPU of NVIDIA's Blackwell architecture (compute capability 120)." << std::endl;
return 0;
examples/80_blackwell_geforce_sparse_gemm/80b_blackwell_geforce_nvfp4_nvfp4_sparse_gemm.cu
+6 -12
View File
@@ -111,7 +111,7 @@ using KernelScheduleType = cutlass::gemm::KernelSparseTmaWarpSpecializedNvf4Sm1
using ThreadBlockShape = Shape<_128,_128,_256>; // Threadblock's tile size
using ClusterShape = Shape<_1,_1,_1>; // Shape of the threadblocks in a cluster
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
ArchTag, OperatorClass,
ArchTag, OperatorClass,
ThreadBlockShape, ClusterShape,
cutlass::epilogue::collective::EpilogueTileAuto,
ElementAccumulator, ElementAccumulator,
@@ -186,13 +186,7 @@ cutlass::HostTensor<ElementCompute, cutlass::layout::PackedVectorLayout> block_N
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
template <typename T>
auto make_iterator(T* ptr) {
using namespace cute;
if constexpr (cute::is_subbyte_v<T>) {
return subbyte_iterator<T>(ptr);
}
else {
return ptr;
}
return cute::recast_ptr<T>(ptr);
}
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Testbed utility types
@@ -300,7 +294,7 @@ bool initialize_block(
}
cutlass::reference::host::TensorFillRandomUniform(
view, seed, scope_max, scope_min, 0);
return true;
}
/// Initialize blocks that released to sparse Matrix A and its metadata E
@@ -489,7 +483,7 @@ template <typename Gemm>
int run(Options &options)
{
// Initialization
if(!initialize(options))
if(!initialize(options))
{
std::cerr << " Initialization failed! " << std::endl;
exit(-1);
@@ -551,9 +545,9 @@ int main(int argc, char const **args) {
cudaDeviceProp props;
int current_device_id;
CUDA_CHECK(cudaGetDevice(&current_device_id));
CUDA_CHECK(cudaGetDeviceProperties(&props, current_device_id));
if (!(props.major == 12 && props.minor == 0)) {
std::cerr << "This example requires a GPU of NVIDIA's Blackwell architecture (compute capability 120)." << std::endl;
return 0;