leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 25 Packages Projects Releases Wiki Activity

update 3.8 v2 (#2112)

Browse Source 
* update 3.8 v2

* update 3.8

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
This commit is contained in:
Yujia Zhai
2025-02-19 22:03:14 -05:00
committed by GitHub
co-authored by yuzhai
parent e9627ce55b
commit b84e9802d8
166 changed files with 3986 additions and 4037 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/72_blackwell_narrow_precision_gemm/72a_blackwell_nvfp4_bf16_gemm.cu
+1 -2
View File
@@ -117,11 +117,10 @@ using OperatorClass = cutlass::arch::OpClassBlockScaledTensorOp; // O
// Kernel Perf config
using MmaTileShape = Shape<_256,_256,_256>; // MMA's tile size
using ClusterShape = Shape<_4,_4,_1>; // Shape of the threadblocks in a cluster
using PerSmTileShape_MNK = Shape<_128,_256,_256>; // Threadblock-level tile size
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
ArchTag, OperatorClass,
PerSmTileShape_MNK, ClusterShape,
MmaTileShape, ClusterShape,
cutlass::epilogue::collective::EpilogueTileAuto,
ElementAccumulator, ElementAccumulator,
ElementC, LayoutCTag, AlignmentC,
examples/72_blackwell_narrow_precision_gemm/72b_blackwell_nvfp4_nvfp4_gemm.cu
+1 -2
View File
@@ -121,7 +121,6 @@ using OperatorClass = cutlass::arch::OpClassBlockScaledTensorOp; // O
// Kernel Perf config
using MmaTileShape = Shape<_128,_128,_256>; // MMA's tile size
using ClusterShape = Shape<_1,_1,_1>; // Shape of the threadblocks in a cluster
using PerSmTileShape_MNK = Shape<_128,_128,_256>; // Threadblock-level tile size
constexpr int InputSFVectorSize = 16;
constexpr int OutputSFVectorSize = InputSFVectorSize;
@@ -137,7 +136,7 @@ using FusionOperation = cutlass::epilogue::fusion::LinCombBlockScaleFactor<
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
ArchTag, OperatorClass,
PerSmTileShape_MNK, ClusterShape,
MmaTileShape, ClusterShape,
cutlass::epilogue::collective::EpilogueTileAuto,
ElementAccumulator, ElementAccumulator,
ElementC, LayoutCTag, AlignmentC,
examples/72_blackwell_narrow_precision_gemm/72c_blackwell_mixed_mxfp8_bf16_gemm.cu
+1 -2
View File
@@ -118,11 +118,10 @@ using OperatorClass = cutlass::arch::OpClassBlockScaledTensorOp; // O
// Kernel Perf config
using MmaTileShape = Shape<_256,_256,_256>; // MMA's tile size
using ClusterShape = Shape<_4,_4,_1>; // Shape of the threadblocks in a cluster
using PerSmTileShape_MNK = Shape<_128,_256,_256>; // Threadblock-level tile size
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
ArchTag, OperatorClass,
PerSmTileShape_MNK, ClusterShape,
MmaTileShape, ClusterShape,
cutlass::epilogue::collective::EpilogueTileAuto,
ElementAccumulator, ElementAccumulator,
ElementC, LayoutCTag, AlignmentC,
examples/72_blackwell_narrow_precision_gemm/CMakeLists.txt
+1 -1
View File
@@ -28,7 +28,7 @@
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
if(NOT CUTLASS_NVCC_ARCHS STREQUAL "100")
if (CUTLASS_NVCC_ARCHS MATCHES 100a)
cutlass_example_add_executable(
72a_blackwell_nvfp4_bf16_gemm
72a_blackwell_nvfp4_bf16_gemm.cu