提交历史

作者 SHA1 备注 提交日期
  Tri Dao 40e534a7f6 Implement cache_leftpad 5 月之前
  Tri Dao dca6d89da4 Don't support softcap and dropout at the same time 5 月之前
  Tri Dao 908511b2b6 Split into more .cu files to speed up compilation 5 月之前
  Tri Dao 1d536d7de5 Minor cleanup of softcapping 5 月之前
  Nicolas Patry 8f873cc6ac Implement softcapping. (#1025) 5 月之前
  Nicolas Patry 5bf201966a Fixing argument checking when using `seqlenq_ngroups_swapped`. (#976) 5 月之前
  Grigory Sizov f816dee63c Support unpadded LSE layout (#970) 5 月之前
  Tri Dao 9eb3d099c1 Transpose out when swapping seqlen_q and num_groups 8 月之前
  Driss Guessous 4a73e903da Add in, macrosf for defining __grid_constant__ (#852) 9 月之前
  Grigory Sizov 2a15840f09 Enable paged attention in varlen forward (#831) 9 月之前
  Tri Dao 2406f28805 Enable headdim 256 backward on consumer GPUs (Ampere, Ada) 9 月之前
  Tri Dao d9a5cb291c Fix dv = torch::empty_like(k) for mha_bwd_varlen as well 10 月之前
  Brian Hirsh 2423cca3ad fix backward for when query and key have different contiguity (#818) 10 月之前
  Grigory Sizov 4687936413 Fix Windows build (#816) 10 月之前
  Jeremy Reizenstein 0658e320f6 Preprocessor switches to control functionality (#788) 10 月之前
  Tri Dao 54e80a3829 Implement page KV cache 10 月之前
  Tri Dao ea8a25ca38 Remove configure in bwd kernel launch 10 月之前
  Grigory Sizov af01244ddd Add split-kv and M<->H swap to varlen forward decoding attention (#754) 10 月之前
  Tri Dao 0842ec0da4 Don't dispatch to local if window size >= seqlen_k 11 月之前
  Tri Dao 732654583c Implement deterministic backward (thanks to Meituan) 11 月之前
  Tri Dao 5ab9b3667b Clean up alibi, implement non-causal alibi 1 年之前
  Sanghun Cho e4f726fc44 Support alibi, by Sanghun Cho from Kakao Brain 1 年之前
  Jeremy Reizenstein ce3e7280f8 Allow varlen_fwd to take optional seqused_k (#647) 1 年之前
  Tri Dao db2f80692c Write zero to out / grad if seqlen_q or seqlen_k is zero 1 年之前
  Tri Dao e279bf8ed9 [Gen] Accept cache_batch_idx to index into the KV cache 1 年之前
  Tri Dao 083e8f525f Implement local attention 1 年之前
  Tri Dao 65c234ed90 Don't over-allocate dq_accum in case of varlen 1 年之前
  Tri Dao 2d8ea9a530 Swap seqlen_q and ngroups when seqlen_q=1 (h/t Daniel Haziza) 1 年之前
  Tri Dao 3250ff3d82 Swap seqlen_q, nheads for MQA when seqlen_q=1 for fwd (h/t Daniel H) 1 年之前
  Tri Dao ccbb14f38e Implement rotary embedding in flash_attn_with_kvcache 1 年之前