Commits · 79bd1a2d5d76b7682d1a78f8b6e529cf11236159 · gaoqiong / flash-attention

13 Nov, 2023 1 commit
- [LayerNorm] Implement residual + LayerNorm/RMSNorm in Triton · 79bd1a2d
  Tri Dao authored Nov 13, 2023
  
  79bd1a2d
09 Nov, 2023 1 commit
- Fix typo in RotaryEmbedding forward output type (#666) · 3566596a
  Antony Frolov authored Nov 09, 2023
  
  3566596a
24 Oct, 2023 2 commits
- Bump to v2.3.3 · 83aef842
  Tri Dao authored Oct 24, 2023
  
  83aef842
- [CrossEntropy] Fix triton cross_entropy_loss IMA for >=2B elements · c79de85f
  Tri Dao authored Oct 24, 2023
  
  c79de85f
12 Oct, 2023 1 commit
- Clarify inference README is a placeholder · 02ac572f
  Tri Dao authored Oct 12, 2023
  
  02ac572f
09 Oct, 2023 1 commit
- Bump to v2.3.2 · 7f31e7c1
  Tri Dao authored Oct 08, 2023
  
  7f31e7c1
08 Oct, 2023 1 commit
- Change constexpr int to constexpr static int · 5a834254
  Tri Dao authored Oct 08, 2023
  
  5a834254
05 Oct, 2023 2 commits
- Add change log · 3a9fe7b0
  Tri Dao authored Oct 05, 2023
  
  3a9fe7b0
- Clarify that Windows is not supported right now · aa4fd2d1
  Tri Dao authored Oct 05, 2023
  
  aa4fd2d1
04 Oct, 2023 2 commits
- [CI] Use official Pytorch 2.1, add CUDA 11.8 for Pytorch 2.1 · 5e525a8d
  Tri Dao authored Oct 03, 2023
  
  5e525a8d
- Bump to v2.3.1 · 21c3b0d8
  Tri Dao authored Oct 03, 2023
  
  21c3b0d8
03 Oct, 2023 1 commit
- [Gen] Accept cache_batch_idx to index into the KV cache · e279bf8e
  Tri Dao authored Oct 03, 2023
  
  e279bf8e
27 Sep, 2023 1 commit
- Bump to v2.3.0 · 601b4dc4
  Tri Dao authored Sep 26, 2023
  
  601b4dc4
26 Sep, 2023 1 commit
- Implement local attention · 083e8f52
  Tri Dao authored Sep 24, 2023
```
Co-authored-by: Timothee Lacroix <t@mistral.ai>
```
  083e8f52
25 Sep, 2023 1 commit
- Fix NameError and typo in ApplyRotaryEmbQKV_ (#569) · 4c8ff915
  Katherine Crowson authored Sep 25, 2023
  
  4c8ff915
24 Sep, 2023 4 commits
- Bump to v2.2.5 · 0a1d03c7
  Tri Dao authored Sep 24, 2023
  
  0a1d03c7
- Switch cutlass to newer commit to avoid compilation warning · 812cb1c9
  Tri Dao authored Sep 24, 2023
  
  812cb1c9
- Don't over-allocate dq_accum in case of varlen · 65c234ed
  Tri Dao authored Sep 24, 2023
  
  65c234ed
- Reduce number of templates for headdim > 128 · 1879e089
  Tri Dao authored Sep 23, 2023
  
  1879e089
22 Sep, 2023 2 commits
- Add placeholder for inference example · dd9a6fa4
  Tri Dao authored Sep 22, 2023
  
  dd9a6fa4
- Re-enable compilation for Hopper · bff31471
  Tri Dao authored Sep 21, 2023
  
  bff31471
21 Sep, 2023 4 commits
- Fix E1136 (#563) · 187c2a06
  Yuchao Dai authored Sep 22, 2023
  
  187c2a06
- Bump to v2.2.4 · 229080b9
  Tri Dao authored Sep 20, 2023
  
  229080b9
- Swap seqlen_q and ngroups when seqlen_q=1 (h/t Daniel Haziza) · 2d8ea9a5
  Tri Dao authored Sep 20, 2023
  
  2d8ea9a5
- [Llama] Fix some tests, add tests for Llama 2 and CodeLlama · 0705d271
  Tri Dao authored Sep 20, 2023
  
  0705d271
20 Sep, 2023 3 commits
- [Gen] Simplify decode_speculative · e0fbaa70
  Tri Dao authored Sep 18, 2023
  
  e0fbaa70
- [Gen] Rename max_sequence_len->max_seqlen, sequence_len_offset->seqlen_offset · e6a80264
  Tri Dao authored Sep 18, 2023
  
  e6a80264
- Fix Llama GQA/MQA (#546) · 42832575
  Kevin Hu authored Sep 19, 2023
```
* Fix llama MQA

* Fix permute shape

* Update llama.py
```
  42832575
18 Sep, 2023 5 commits
- [Gen] Don't use ft_attention, use flash_attn_with_kvcache instead · dfe29f5e
  Tri Dao authored Sep 18, 2023
  
  dfe29f5e
- Swap seqlen_q, nheads for MQA when seqlen_q=1 for fwd (h/t Daniel H) · 3250ff3d
  Tri Dao authored Sep 18, 2023
  
  3250ff3d
- Remove template for (IsEvenMN=T, IsEvenK=F) to speed up compilation · 43617dea
  Tri Dao authored Sep 18, 2023
  
  43617dea
- [Minor] add nvcc note on bare_metal_version `RuntimeError` (#552) · fa3ddcba
  Federico Berto authored Sep 19, 2023
```
* Add nvcc note on bare_metal_version `RuntimeError`

* Run Black formatting
```
  fa3ddcba
- Don't compile for Pytorch 2.1 on CUDA 12.1 due to nvcc segfaults · 799f56fa
  Tri Dao authored Sep 17, 2023
  
  799f56fa
17 Sep, 2023 1 commit
- Set block size to 64 x 64 for kvcache to avoid nvcc segfaults · c984208d
  Tri Dao authored Sep 17, 2023
  
  c984208d
16 Sep, 2023 3 commits
- Bump to v2.2.3 · 8c8b4d36
  Tri Dao authored Sep 16, 2023
  
  8c8b4d36
- Implement rotary embedding in flash_attn_with_kvcache · ccbb14f3
  Tri Dao authored Sep 16, 2023
  
  ccbb14f3
- [CE] Implement CrossEntropyLoss in Triton · 5400fdc4
  Tri Dao authored Sep 15, 2023
  
  5400fdc4
13 Sep, 2023 2 commits
- Simplify the implementation of KVcache attn by appending KV first · 56b7fc6e
  Tri Dao authored Sep 13, 2023
  
  56b7fc6e
- Add tests for Pythia, GPT-JT, and RedPajama models · d0032700
  Tri Dao authored Sep 13, 2023
  
  d0032700
12 Sep, 2023 1 commit
- Remove some unused headers · bb9beb36
  Tri Dao authored Sep 12, 2023
  
  bb9beb36