Commits · 5ab9b3667bc39a0063ed588be63fe8989d9da080 · gaoqiong / flash-attention

22 Dec, 2023 1 commit
- Clean up alibi, implement non-causal alibi · 5ab9b366
  Tri Dao authored Dec 21, 2023
  
  5ab9b366
03 Oct, 2023 1 commit
- [Gen] Accept cache_batch_idx to index into the KV cache · e279bf8e
  Tri Dao authored Oct 03, 2023
  
  e279bf8e
26 Sep, 2023 1 commit
- Implement local attention · 083e8f52
  Tri Dao authored Sep 24, 2023
```
Co-authored-by: Timothee Lacroix <t@mistral.ai>
```
  083e8f52
24 Sep, 2023 1 commit
- Don't over-allocate dq_accum in case of varlen · 65c234ed
  Tri Dao authored Sep 24, 2023
  
  65c234ed
21 Sep, 2023 1 commit
- Swap seqlen_q and ngroups when seqlen_q=1 (h/t Daniel Haziza) · 2d8ea9a5
  Tri Dao authored Sep 20, 2023
  
  2d8ea9a5
18 Sep, 2023 1 commit
- Swap seqlen_q, nheads for MQA when seqlen_q=1 for fwd (h/t Daniel H) · 3250ff3d
  Tri Dao authored Sep 18, 2023
  
  3250ff3d
16 Sep, 2023 1 commit
- Implement rotary embedding in flash_attn_with_kvcache · ccbb14f3
  Tri Dao authored Sep 16, 2023
  
  ccbb14f3
13 Sep, 2023 1 commit
- Simplify the implementation of KVcache attn by appending KV first · 56b7fc6e
  Tri Dao authored Sep 13, 2023
  
  56b7fc6e
04 Sep, 2023 2 commits
- Implement flash_attn_with_kvcache · 37c6e054
  Tri Dao authored Sep 04, 2023
  
  37c6e054
- Require CUDA 11.6+, clean up setup.py · 0c04943f
  Tri Dao authored Sep 03, 2023
  
  0c04943f
29 Aug, 2023 1 commit
- Implement splitKV attention · b1fbbd83
  Tri Dao authored Aug 29, 2023
  
  b1fbbd83
25 Aug, 2023 1 commit
- Change causal mask to be aligned to bottom-right instead of top-left · 9e5e8bc9
  Tri Dao authored Aug 21, 2023
  
  9e5e8bc9
19 Aug, 2023 1 commit
- Run isort and black on test files · 0e8c46ae
  Tri Dao authored Aug 18, 2023
  
  0e8c46ae
16 Aug, 2023 1 commit
- Fix Bwd NaN for varlen when seqlen_q >> seqlen_k and causal · c65b5106
  Tri Dao authored Aug 16, 2023
  
  c65b5106
13 Aug, 2023 1 commit
- Update to Cutlass 3.1 · 3524e13c
  Tri Dao authored Aug 13, 2023
  
  3524e13c
01 Aug, 2023 2 commits
- Fix race condition in bwd (overwriting sK) · 1c41d2b0
  Tri Dao authored Aug 01, 2023
  
  1c41d2b0
- Fix masking of bwd when seqlen is not divisible by 128 · a4f148b6
  Tri Dao authored Jul 31, 2023
  
  a4f148b6
17 Jul, 2023 1 commit
- FlashAttention-2 release · 4f285b35
  Tri Dao authored Jul 17, 2023
  
  4f285b35
13 Nov, 2022 2 commits
- Skip flash_attn_split test · a8fec99a
  Tri Dao authored Nov 13, 2022
  
  a8fec99a
- Don't enforce bitwise consistency for dq in race condition test · 9d3116ad
  Tri Dao authored Nov 13, 2022
```
Since we could be parallelizing over seqlen_k
```
  9d3116ad
09 Nov, 2022 1 commit
- Fix out-of-bound memory read · 6998e0ec
  Tri Dao authored Nov 09, 2022
  
  6998e0ec
06 Nov, 2022 1 commit
- Fix pipelining bug in Triton bwd with bias_type=matrix · 74797571
  Tri Dao authored Nov 06, 2022
  
  74797571
05 Nov, 2022 1 commit
- Parallelize CUDA bwd along seqlen_k instead of seqlen_q · 55778193
  Tri Dao authored Nov 05, 2022
```
This is faster since we only need to do atomic adds on dq, instead of atomic
adds on both dk and dv.
```
  55778193
04 Nov, 2022 2 commits
- Fix race condition in Triton bwd when there's bias · ff78ea41
  Tri Dao authored Nov 04, 2022
  
  ff78ea41
- Implement attention bias for Triton version · 86862cfd
  Tri Dao authored Nov 04, 2022
  
  86862cfd
02 Nov, 2022 1 commit
- Fix race condition in Triton bwd for non-po2 headdims · aacc10fb
  Tri Dao authored Nov 02, 2022
  
  aacc10fb
01 Nov, 2022 1 commit
- Avoid memcpy in the Triton bwd · 1fb12afd
  Tri Dao authored Nov 01, 2022
  
  1fb12afd
31 Oct, 2022 8 commits
- Fix race condition in Triton fwd · 9b0bc978
  Tri Dao authored Oct 31, 2022
  
  9b0bc978
- Add debug_barrier for all headdims in Triton bwd · 4f81aff4
  Tri Dao authored Oct 31, 2022
  
  4f81aff4
- [WIP] Support all head dimensions up to 128 in the Triton bwd · e78d509c
  Tri Dao authored Oct 31, 2022
```
WIP because there seems to be some race conditions for head dimensions other
than 16, 32, 64, 128.
```
  e78d509c
- Support all head dimensions up to 128 in the Triton fwd · 008951f1
  Tri Dao authored Oct 30, 2022
  
  008951f1
- Support arbitrary seqlens (both q & k) in Triton bwd · b910bf14
  Tri Dao authored Oct 30, 2022
  
  b910bf14
- Support arbitrary seqlen_k in Triton bwd · dc554693
  Tri Dao authored Oct 30, 2022
  
  dc554693
- Fix Triton fwd to support seqlen not multiples of 128 · d11341fd
  Tri Dao authored Oct 30, 2022
  
  d11341fd
- Implement FlashAttention in Triton · b0c0db81
  Tri Dao authored Oct 30, 2022
  
  b0c0db81
24 Oct, 2022 1 commit
- Support all head dims that are multiples of 8, up to 128 · 46fd2a20
  Tri Dao authored Oct 24, 2022
  
  46fd2a20
23 Oct, 2022 1 commit
- Split bwd on the seqlen_q dimension · a5a8806d
  Tri Dao authored Oct 23, 2022
  
  a5a8806d
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
16 Oct, 2022 1 commit
- Fix #54: set device for multi-GPU case · 52fb4b72
  Tri Dao authored Oct 16, 2022
  
  52fb4b72
14 Oct, 2022 1 commit
- Implement attention kernel that splits the batch into two · 5badfb78
  Tri Dao authored Oct 13, 2022
  
  5badfb78