squall/flash-attention
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
b17c6fe235
flash-attention/csrc/flash_attn/src
History
Tri Dao b17c6fe235 Reduce smem usage for Q and dO in the backward pass 
From 4KB per buffer to 2KB per buffer. This saves us 8KB of smem (each Q and dO
have 2 buffers)
2022-06-03 16:59:11 -07:00
..
fmha Reduce smem usage for Q and dO in the backward pass 2022-06-03 16:59:11 -07:00
.DS_Store Rename, add benchmarking script 2022-05-26 13:57:38 -07:00
fmha_block_dgrad_fp16_kernel_loop.sm80.cu Rename, add benchmarking script 2022-05-26 13:57:38 -07:00
fmha_block_dgrad_kernel_1xN_loop.h Use Cutlass gemm as WarpMma 2022-06-02 10:33:32 -07:00
fmha_block_fprop_fp16_kernel.sm80.cu Rename, add benchmarking script 2022-05-26 13:57:38 -07:00
fmha_block_fprop_kernel_1xN.h Use Cutlass gemm as WarpMma 2022-06-02 10:33:32 -07:00
fmha_blockmask.h Rename, add benchmarking script 2022-05-26 13:57:38 -07:00
fmha_dgrad_fp16_kernel_loop.sm80.cu Support Turing mma instructions 2022-06-03 16:58:44 -07:00
fmha_dgrad_kernel_1xN_loop.h Use Cutlass gemm as WarpMma 2022-06-02 10:33:32 -07:00
fmha_fprop_fp16_kernel.sm80.cu Support Turing mma instructions 2022-06-03 16:58:44 -07:00
fmha_fprop_kernel_1xN.h Use Cutlass gemm as WarpMma 2022-06-02 10:33:32 -07:00
fmha_kernel.h Rename, add benchmarking script 2022-05-26 13:57:38 -07:00
fmha_utils.h Rename, add benchmarking script 2022-05-26 13:57:38 -07:00
fmha.h Rename, add benchmarking script 2022-05-26 13:57:38 -07:00
philox.cuh Rename, add benchmarking script 2022-05-26 13:57:38 -07:00