diff options
Diffstat (limited to 'candle-flash-attn/kernels/flash_fwd_hdim224_fp16_causal_sm80.cu')
| -rw-r--r-- | candle-flash-attn/kernels/flash_fwd_hdim224_fp16_causal_sm80.cu | 10 |
1 files changed, 10 insertions, 0 deletions
diff --git a/candle-flash-attn/kernels/flash_fwd_hdim224_fp16_causal_sm80.cu b/candle-flash-attn/kernels/flash_fwd_hdim224_fp16_causal_sm80.cu new file mode 100644 index 00000000..f01dad09 --- /dev/null +++ b/candle-flash-attn/kernels/flash_fwd_hdim224_fp16_causal_sm80.cu @@ -0,0 +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, 224, true>(Flash_fwd_params ¶ms, cudaStream_t stream) { + run_mha_fwd_hdim224<cutlass::half_t, true>(params, stream); +} |