candle-flash-attn/kernels/flash_fwd_hdim64_bf16_sm80.cu


1
2
3
4
5
6
7
8
9
10

// Copyright (c) 2023, Tri Dao.
// Splitting the different head dimensions to different files to speed up compilation.
// This file is auto-generated. See "generate_kernels.py"

#include "flash_fwd_launch_template.h"

template<>
void run_mha_fwd_<cutlass::bfloat16_t, 64>(Flash_fwd_params &params, cudaStream_t stream) {
    run_mha_fwd_hdim64<cutlass::bfloat16_t>(params, stream);
}