value Class — pytorch Architecture

Source Code

aten/src/ATen/native/cuda/cutlass_extensions/gemm/kernel/mixed_gemm_B_layout.h lines 41–47

template<typename Arch>
struct LayoutDetailsB<half_t, Arch, typename platform::enable_if<Arch::kMinComputeCapability >= 75>::type> {
    static constexpr int ThreadblockK      = 64;
    using Layout                           = layout::RowMajor;
    static constexpr int ElementsPerAccess = 128 / cutlass::sizeof_bits<half_t>::value;
    using Operator                         = cutlass::arch::OpMultiplyAdd;
};