Commits · 23b77c8148f4ee87cecc8a577e92fe685ff73316 · gaoqiong / flash-attention

"coco_retinanet_gpu.yaml" did not exist on "eeb371d41da1be272a9c00be466bf8b1ecee790b"

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
16 Sep, 2023 1 commit
- Implement rotary embedding in flash_attn_with_kvcache · ccbb14f3
  Tri Dao authored Sep 16, 2023
  
  ccbb14f3
11 Sep, 2023 1 commit
- Swap seqlen_q and nheads for MQA to speed it up (h/t Daniel Haziza) · ee77b931
  Tri Dao authored Sep 10, 2023
  
  ee77b931
05 Sep, 2023 1 commit
- Support cache_seqlens being integer · fd20f16a
  Tri Dao authored Sep 05, 2023
  
  fd20f16a
04 Sep, 2023 1 commit
- Implement flash_attn_with_kvcache · 37c6e054
  Tri Dao authored Sep 04, 2023
  
  37c6e054
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
20 Aug, 2023 1 commit
- Import torch before flash_attn_2_cuda · d431f167
  Tri Dao authored Aug 19, 2023
  
  d431f167
18 Aug, 2023 1 commit
- Run isort and black on python files · f1a73d07
  Tri Dao authored Aug 18, 2023
  
  f1a73d07
01 Aug, 2023 1 commit
- [Docs] Fix docstring about Q nheads being divisible by KV nheads · 8f4cd4c1
  Tri Dao authored Jul 31, 2023
  
  8f4cd4c1
28 Jul, 2023 1 commit
- [Docs] Fix mention of MQA/GQA in qkvpacked functions · 840f7925
  Tri Dao authored Jul 28, 2023
  
  840f7925
27 Jul, 2023 1 commit

Kirthi Shankar Sivamani authored Jul 27, 2023



* Add RNG state to kernel launch params
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Save seed and offset for backward
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Single thread write to global mem
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* compute_dq_dk_dv_1colblock get seed and offset from launch params
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* compute_dq_dk_dv_1rowblock get seed and offset from launch params
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Change forward c++ APIs to save RNG state for backward
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Change backward c++ APIs to set RNG state for bprop launcher
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Bug fixes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Python side API changes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Bug fix; only save seeds instead of full offset
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Account for 3D grid size
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

a03f6f8e

18 Jul, 2023 1 commit
- Make sure dout is contiguous · b4cc152e
  Tri Dao authored Jul 17, 2023
  
  b4cc152e
17 Jul, 2023 1 commit
- FlashAttention-2 release · 4f285b35
  Tri Dao authored Jul 17, 2023
  
  4f285b35
03 Jul, 2023 1 commit
- [Doc] Change total -> total_q · e8a0b4ac
  Tri Dao authored Jul 02, 2023
  
  e8a0b4ac
13 Apr, 2023 1 commit
- Handle FlashAttnQKVPackedSplitFunc by making rng_state optional in backward · 7d25a4ec
  Kirthi Shankar Sivamani authored Apr 13, 2023
```
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
```
  7d25a4ec
12 Apr, 2023 1 commit
- Support CUDA graph capture · 31018c5f
  Kirthi Shankar Sivamani authored Apr 12, 2023
```
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
```
  31018c5f
31 Mar, 2023 1 commit
- Add option for deterministic execution · b6aa059b
  Kirthi Shankar Sivamani authored Mar 30, 2023
  
  b6aa059b
13 Dec, 2022 1 commit
- Fix the case when dout is not contiguous · 88c4e5db
  Tri Dao authored Dec 13, 2022
  
  88c4e5db
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
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 2 commits
- Split fwd on the seqlen_q dimension · a44f48df
  Tri Dao authored Oct 21, 2022
  
  a44f48df
- 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
14 Oct, 2022 2 commits
- Fix QKV interface to allocate output in Python · 1b9facac
  Tri Dao authored Oct 14, 2022
  
  1b9facac
- Implement attention kernel that splits the batch into two · 5badfb78
  Tri Dao authored Oct 13, 2022
  
  5badfb78
04 Jul, 2022 2 commits
- Do P * dP (pointwise) in the bwd in fp32 instead of fp16 · a5559a0e
  Tri Dao authored Jul 03, 2022
  
  a5559a0e
- Implement cross attention · 6c3a8c65
  Tri Dao authored Jun 30, 2022
  
  6c3a8c65
02 Jun, 2022 1 commit
- Rename src -> flash_attn · 5a61cb77
  Tri Dao authored Jun 01, 2022
  
  5a61cb77
29 May, 2022 1 commit
- Reorganize directories, add banner figure · 67c37795
  Tri Dao authored May 29, 2022
  
  67c37795
26 May, 2022 1 commit
- Rename, add benchmarking script · 9dbc491a
  Tri Dao authored May 26, 2022
  
  9dbc491a
20 May, 2022 1 commit
- First release · 1fcbe6f0
  Tri Dao authored May 20, 2022
  
  1fcbe6f0