// 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, 96, true>(Flash_fwd_params& params, cudaStream_t stream) { run_mha_fwd_hdim96<cutlass::bfloat16_t, true>(params, stream); }