Commit History

Author SHA1 Message Date
  Antoni Viros 83e41b3ca4 Add custom ops for compatibility with PT Compile (#1139) 3 months ago
  Tri Dao 65f723bb9a Split bwd into more .cu files to speed up compilation 5 months ago
  Tri Dao 751c762c9c Don't specialize for hdim 224 to speed up compilation 5 months ago
  Driss Guessous 1c275eb070 Fix ima for split-kv kernel (#1085) 5 months ago
  Jorge António 4df62e1440 catch typo (#1058) 5 months ago
  Tri Dao 40e534a7f6 Implement cache_leftpad 5 months ago
  Tri Dao dca6d89da4 Don't support softcap and dropout at the same time 6 months ago
  Tri Dao 908511b2b6 Split into more .cu files to speed up compilation 6 months ago
  Tri Dao 1d536d7de5 Minor cleanup of softcapping 6 months ago
  Nicolas Patry 8f873cc6ac Implement softcapping. (#1025) 6 months ago
  Nicolas Patry 5bf201966a Fixing argument checking when using `seqlenq_ngroups_swapped`. (#976) 6 months ago
  Grigory Sizov f816dee63c Support unpadded LSE layout (#970) 6 months ago
  Tri Dao 9eb3d099c1 Transpose out when swapping seqlen_q and num_groups 9 months ago
  Driss Guessous 4a73e903da Add in, macrosf for defining __grid_constant__ (#852) 9 months ago
  Grigory Sizov 2a15840f09 Enable paged attention in varlen forward (#831) 9 months ago
  Tri Dao 2406f28805 Enable headdim 256 backward on consumer GPUs (Ampere, Ada) 10 months ago
  Tri Dao d9a5cb291c Fix dv = torch::empty_like(k) for mha_bwd_varlen as well 11 months ago
  Brian Hirsh 2423cca3ad fix backward for when query and key have different contiguity (#818) 11 months ago
  Grigory Sizov 4687936413 Fix Windows build (#816) 11 months ago
  Jeremy Reizenstein 0658e320f6 Preprocessor switches to control functionality (#788) 11 months ago
  Tri Dao 54e80a3829 Implement page KV cache 11 months ago
  Tri Dao ea8a25ca38 Remove configure in bwd kernel launch 11 months ago
  Grigory Sizov af01244ddd Add split-kv and M<->H swap to varlen forward decoding attention (#754) 11 months ago
  Tri Dao 0842ec0da4 Don't dispatch to local if window size >= seqlen_k 1 year ago
  Tri Dao 732654583c Implement deterministic backward (thanks to Meituan) 1 year ago
  Tri Dao 5ab9b3667b Clean up alibi, implement non-causal alibi 1 year ago
  Sanghun Cho e4f726fc44 Support alibi, by Sanghun Cho from Kakao Brain 1 year ago
  Jeremy Reizenstein ce3e7280f8 Allow varlen_fwd to take optional seqused_k (#647) 1 year ago
  Tri Dao db2f80692c Write zero to out / grad if seqlen_q or seqlen_k is zero 1 year ago
  Tri Dao e279bf8ed9 [Gen] Accept cache_batch_idx to index into the KV cache 1 year ago