value Class — pytorch Architecture

Source Code

aten/src/ATen/native/transformers/cuda/mem_eff_attention/gemm_kernel_utils.h lines 128–141

template <typename ArchTag, typename scalar_t>
struct DefaultGemmType<
    ArchTag,
    scalar_t,
    typename cutlass::platform::enable_if<
        ArchTag::kMinComputeCapability >= 75 &&
        cutlass::sizeof_bits<scalar_t>::value == 16>::type> {
  static constexpr int ThreadK = 32;
  static constexpr int WarpK = 32;
  static constexpr int kMinimumAlignment = 4;
  using OpClass = cutlass::arch::OpClassTensorOp;
  using InstructionShape = cutlass::gemm::GemmShape<16, 8, 8>;
  using Operator = cutlass::arch::OpMultiplyAdd;
};