// Copyright (c) 2024, Tri Dao. | |
// Splitting the different head dimensions to different files to speed up compilation. | |
template void run_mha_fwd_combine_<float, float, 64>(Flash_fwd_params ¶ms, cudaStream_t stream, bool enable_pdl); | |
template void run_mha_fwd_combine_<float, float, 128>(Flash_fwd_params ¶ms, cudaStream_t stream, bool enable_pdl); | |
template void run_mha_fwd_combine_<cutlass::half_t, float, 64>(Flash_fwd_params ¶ms, cudaStream_t stream, bool enable_pdl); | |
template void run_mha_fwd_combine_<cutlass::half_t, float, 128>(Flash_fwd_params ¶ms, cudaStream_t stream, bool enable_pdl); | |
template void run_mha_fwd_combine_<cutlass::bfloat16_t, float, 64>(Flash_fwd_params ¶ms, cudaStream_t stream, bool enable_pdl); | |
template void run_mha_fwd_combine_<cutlass::bfloat16_t, float, 128>(Flash_fwd_params ¶ms, cudaStream_t stream, bool enable_pdl); | |