DefaultGemmType Class — pytorch Architecture

Source Code

aten/src/ATen/native/transformers/cuda/mem_eff_attention/gemm_kernel_utils.h lines 144–152

template <>
struct DefaultGemmType<cutlass::arch::Sm70, cutlass::half_t, void> {
  static constexpr int ThreadK = 32;
  static constexpr int WarpK = 32;
  static constexpr int kMinimumAlignment = 2;
  using OpClass = cutlass::arch::OpClassTensorOp;
  using InstructionShape = cutlass::gemm::GemmShape<8, 8, 4>;
  using Operator = cutlass::arch::OpMultiplyAdd;
};