1
0

Коммит түүх

Эзэн SHA1 Мессеж Огноо
  Antoni Viros 83e41b3ca4 Add custom ops for compatibility with PT Compile (#1139) 2 сар өмнө
  youkaichao ef3e358a25 remove lambda (#1056) 4 сар өмнө
  Tri Dao 898dd4bbf2 Pass seqused_k to _flash_attn_varlen_forward 5 сар өмнө
  Tri Dao 40e534a7f6 Implement cache_leftpad 5 сар өмнө
  Tri Dao 81e01efd4b More typo fixes 5 сар өмнө
  Tri Dao 72e27c6320 Fix typo with softcapping 5 сар өмнө
  Phil Wang f4628b43ec missing commas and backwards return arguments (#1032) 5 сар өмнө
  Nicolas Patry 8f873cc6ac Implement softcapping. (#1025) 5 сар өмнө
  Jianwei Dong 4e8d60069f Add the return_softmax_lse parameter to the flash_attn_with_kvcache function to allow returning the logsumexp of the attention scores. (#989) 5 сар өмнө
  Grigory Sizov f816dee63c Support unpadded LSE layout (#970) 5 сар өмнө
  Grigory Sizov 2a15840f09 Enable paged attention in varlen forward (#831) 9 сар өмнө
  Tao He 204c3c6d1b Fixes an error in comment (#785) 10 сар өмнө
  Tri Dao 54e80a3829 Implement page KV cache 10 сар өмнө
  Tri Dao a7b66ae25a Simplify writing softmax to gmem 11 сар өмнө
  Tri Dao 732654583c Implement deterministic backward (thanks to Meituan) 11 сар өмнө
  Tri Dao 5ab9b3667b Clean up alibi, implement non-causal alibi 1 жил өмнө
  Tri Dao bc28eacc60 Format flash_attn_interface.py 1 жил өмнө
  Sanghun Cho e4f726fc44 Support alibi, by Sanghun Cho from Kakao Brain 1 жил өмнө
  Tri Dao d4a7c8ffbb [CI] Only compile for CUDA 11.8 & 12.2, MAX_JOBS=2,add torch-nightly 1 жил өмнө
  Jeremy Reizenstein ce3e7280f8 Allow varlen_fwd to take optional seqused_k (#647) 1 жил өмнө
  Tri Dao e279bf8ed9 [Gen] Accept cache_batch_idx to index into the KV cache 1 жил өмнө
  Tri Dao 083e8f525f Implement local attention 1 жил өмнө
  Tri Dao ccbb14f38e Implement rotary embedding in flash_attn_with_kvcache 1 жил өмнө
  Tri Dao ee77b931b9 Swap seqlen_q and nheads for MQA to speed it up (h/t Daniel Haziza) 1 жил өмнө
  Tri Dao fd20f16a4e Support cache_seqlens being integer 1 жил өмнө
  Tri Dao 37c6e05406 Implement flash_attn_with_kvcache 1 жил өмнө
  Tri Dao 9e5e8bc91e Change causal mask to be aligned to bottom-right instead of top-left 1 жил өмнө
  Tri Dao d431f16751 Import torch before flash_attn_2_cuda 1 жил өмнө
  Tri Dao f1a73d0740 Run isort and black on python files 1 жил өмнө
  Tri Dao 8f4cd4c16b [Docs] Fix docstring about Q nheads being divisible by KV nheads 1 жил өмнө