mirror of
https://github.com/huggingface/candle.git
synced 2025-06-16 02:38:10 +00:00
2ce5f12513
* Again set a few extra params. * Use the appropriate kernel sizes. * Add all the kernel sizes. * Parallel compiling. * Reduce the amount of parallelism. * Add the missing kernel. * Fix a typo. * Remove bf16 support for now.
19 lines
777 B
Plaintext
19 lines
777 B
Plaintext
// Copyright (c) 2023, Tri Dao.
|
|
|
|
// Splitting the different head dimensions to different files to speed up compilation.
|
|
|
|
#include "flash_fwd_launch_template.h"
|
|
|
|
// template<>
|
|
// void run_mha_fwd_<cutlass::bfloat16_t, 64>(Flash_fwd_params ¶ms, cudaStream_t stream) {
|
|
// using elem_type = cutlass::bfloat16_t;
|
|
// if (params.p_dropout == 1.f) {
|
|
// run_flash_fwd<Flash_fwd_kernel_traits<64, 128, 64, 4, true, false, elem_type>, false>(params, stream);
|
|
// } else {
|
|
// run_flash_fwd<Flash_fwd_kernel_traits<64, 128, 64, 4, false, false, elem_type>, true>(params, stream);
|
|
// }
|
|
// }
|
|
template<>
|
|
void run_mha_fwd_<cutlass::bfloat16_t, 64>(Flash_fwd_params ¶ms, cudaStream_t stream) {
|
|
run_mha_fwd_hdim64<cutlass::bfloat16_t>(params, stream);
|
|
} |