diff options
Diffstat (limited to 'candle-flash-attn/kernels/flash_fwd_hdim32_fp16_sm80.cu')
| -rw-r--r-- | candle-flash-attn/kernels/flash_fwd_hdim32_fp16_sm80.cu | 17 |
1 files changed, 2 insertions, 15 deletions
diff --git a/candle-flash-attn/kernels/flash_fwd_hdim32_fp16_sm80.cu b/candle-flash-attn/kernels/flash_fwd_hdim32_fp16_sm80.cu index 91e6331e..8dbf8b94 100644 --- a/candle-flash-attn/kernels/flash_fwd_hdim32_fp16_sm80.cu +++ b/candle-flash-attn/kernels/flash_fwd_hdim32_fp16_sm80.cu @@ -1,23 +1,10 @@ // 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::half_t, 32>(Flash_fwd_params ¶ms, cudaStream_t stream) { -// using elem_type = cutlass::half_t; -// BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] { -// run_flash_fwd<Flash_fwd_kernel_traits<32, 128, 128, 4, false, false, elem_type>, Is_dropout>(params, stream); -// // For dropout there might be a lot of register spilling? -// // These two are very slow due to register spilling -// // run_flash_fwd<Flash_fwd_kernel_traits<32, 256, 128, 4, false, elem_type>>(params, stream); -// // run_flash_fwd<Flash_fwd_kernel_traits<32, 128, 256, 4, false, elem_type>>(params, stream); -// // This one is slightly slower -// // run_flash_fwd<Flash_fwd_kernel_traits<32, 256, 64, 4, false, elem_type>>(params, stream); -// }); -// } template<> void run_mha_fwd_<cutlass::half_t, 32>(Flash_fwd_params ¶ms, cudaStream_t stream) { run_mha_fwd_hdim32<cutlass::half_t>(params, stream); -}
\ No newline at end of file +} |