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

v3.8.0 update (#2082)

Browse Source 
* 3.8 update

* fix Markus' name

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
This commit is contained in:
Yujia Zhai
2025-02-06 21:33:40 -05:00
committed by GitHub
co-authored by yuzhai
parent affd1b693d
commit 833f6990e0
168 changed files with 24945 additions and 3436 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/unit/gemm/device/gemm_testbed_3x_ptr_array.hpp
+1 -3
View File
@@ -1119,7 +1119,6 @@ struct HostCollectiveEpilogue {
std::vector<cutlass::HostTensor<ElementSFD, LayoutTagD>> tensors_SFD;
std::vector<cutlass::HostTensor<ElementSFD, LayoutTagD>> references_SFD;
cutlass::DeviceAllocation<ElementSFD *> device_tensors_SFD;
using ElementCompute = typename FusionOp::ElementCompute;
using ElementScalar = typename FusionOp::ElementScalar;
@@ -2205,8 +2204,7 @@ bool TestSmall(double alpha = 1.0, double beta = 1.0,
static constexpr bool IsF8F6F4 = cutlass::gemm::collective::detail::is_sm100_mma_f8f6f4<TiledMma, ElementA, ElementB>();
alignment_bits = cutlass::detail::get_input_alignment_bits<ElementA, IsF8F6F4>();
// For fp4 and fp6 mx kernels, the min alignment_input is 128 elements, so we don't need to add alignment_input in test problem sizes.
// For fp4 and fp6 QMMA kernels, the min alignment_input is 128 elements, so we don't need to add alignment_input in test problem sizes.
int alignment_input = (alignment_bits / cute::sizeof_bits<ElementA>::value == 128) ? 0 : (alignment_bits / cute::sizeof_bits<ElementA>::value);