Commits · 1aa6d7d9b60bf8fbb5584f057934bdee15ed33fe · gaoqiong / flash-attention

21 Oct, 2022 1 commit
- Rework dropout to decouple forward and backward · 1aa6d7d9
  Tri Dao authored Oct 18, 2022
```
They don't have to have the same block size, number of threads, etc.
```
  1aa6d7d9
10 Jul, 2022 3 commits
- Implement for bf16 · de19de7a
  Tri Dao authored Jul 09, 2022
  
  de19de7a
- Refactor gemm_cl to template on either __half or __nv_bfloat16 · 6a77a6da
  Tri Dao authored Jul 08, 2022
  
  6a77a6da
- Refactor to template on __half, implement bf16 util functions · e518a4b3
  Tri Dao authored Jul 08, 2022
  
  e518a4b3
04 Jul, 2022 1 commit
- Implement cross attention · 6c3a8c65
  Tri Dao authored Jun 30, 2022
  
  6c3a8c65
30 Jun, 2022 1 commit
- Support batch size > 64K by swapping grid.x and grid.y · f66603cb
  Tri Dao authored Jun 29, 2022
  
  f66603cb
25 Jun, 2022 1 commit
- Bug fix: wrong smem_o write pointer for d=16 · eeca63a7
  Tri Dao authored Jun 25, 2022
  
  eeca63a7
12 Jun, 2022 3 commits
- Refactor Gmem code to store q, k, v pointers separately · 5d07483b
  Tri Dao authored Jun 12, 2022
  
  5d07483b
- Implement bwd for head dim 128 · d3e64409
  Tri Dao authored Jun 11, 2022
  
  d3e64409
- Implement fwd for head dim 128 · 0d854692
  Tri Dao authored Jun 05, 2022
  
  0d854692
03 Jun, 2022 2 commits
- Reduce smem usage for Q and dO in the backward pass · b17c6fe2
  Tri Dao authored Jun 03, 2022
```
From 4KB per buffer to 2KB per buffer. This saves us 8KB of smem (each Q and dO
have 2 buffers)
```
  b17c6fe2
- Support Turing mma instructions · 2712aa4c
  Tri Dao authored Jun 02, 2022
  
  2712aa4c
02 Jun, 2022 2 commits
- Remove softmax fp16 max · 05087332
  Tri Dao authored Jun 02, 2022
  
  05087332
- Use Cutlass gemm as WarpMma · 14dc326e
  Tri Dao authored Jun 02, 2022
  
  14dc326e
26 May, 2022 1 commit
- Rename, add benchmarking script · 9dbc491a
  Tri Dao authored May 26, 2022
  
  9dbc491a