Commit History

Автор SHA1 Съобщение Дата
  Driss Guessous bc482cbf91 Add a macro for namespace (#1419) преди 4 седмици
  Lu Fang 74aed78373 Replace c10::optional with std::optional in flash_attn преди 4 седмици
  Kirthi Shankar Sivamani bbb21d6533 Generalize arch checks for A100 and above (#1427) преди 1 месец
  Tri Dao 9375ac9322 [CI] Don't include <ATen/cuda/CUDAGraphsUtils.cuh> преди 2 месеца
  Tri Dao e782d28692 [CI] Change torch #include to make it work with torch 2.1 Philox преди 2 месеца
  Tri Dao 073afd5931 [CI] Use torch 2.6.0.dev20241001, reduce torch #include преди 2 месеца
  sclarkson 1feb711f46 Fix compilation with clang on ARM64 (#1285) преди 2 месеца
  Antoni Viros 83e41b3ca4 Add custom ops for compatibility with PT Compile (#1139) преди 4 месеца
  Tri Dao 65f723bb9a Split bwd into more .cu files to speed up compilation преди 6 месеца
  Tri Dao 751c762c9c Don't specialize for hdim 224 to speed up compilation преди 6 месеца
  Driss Guessous 1c275eb070 Fix ima for split-kv kernel (#1085) преди 6 месеца
  Jorge António 4df62e1440 catch typo (#1058) преди 6 месеца
  Tri Dao 40e534a7f6 Implement cache_leftpad преди 7 месеца
  Tri Dao dca6d89da4 Don't support softcap and dropout at the same time преди 7 месеца
  Tri Dao 908511b2b6 Split into more .cu files to speed up compilation преди 7 месеца
  Tri Dao 1d536d7de5 Minor cleanup of softcapping преди 7 месеца
  Nicolas Patry 8f873cc6ac Implement softcapping. (#1025) преди 7 месеца
  Nicolas Patry 5bf201966a Fixing argument checking when using `seqlenq_ngroups_swapped`. (#976) преди 7 месеца
  Grigory Sizov f816dee63c Support unpadded LSE layout (#970) преди 7 месеца
  Tri Dao 9eb3d099c1 Transpose out when swapping seqlen_q and num_groups преди 10 месеца
  Driss Guessous 4a73e903da Add in, macrosf for defining __grid_constant__ (#852) преди 11 месеца
  Grigory Sizov 2a15840f09 Enable paged attention in varlen forward (#831) преди 11 месеца
  Tri Dao 2406f28805 Enable headdim 256 backward on consumer GPUs (Ampere, Ada) преди 11 месеца
  Tri Dao d9a5cb291c Fix dv = torch::empty_like(k) for mha_bwd_varlen as well преди 1 година
  Brian Hirsh 2423cca3ad fix backward for when query and key have different contiguity (#818) преди 1 година
  Grigory Sizov 4687936413 Fix Windows build (#816) преди 1 година
  Jeremy Reizenstein 0658e320f6 Preprocessor switches to control functionality (#788) преди 1 година
  Tri Dao 54e80a3829 Implement page KV cache преди 1 година
  Tri Dao ea8a25ca38 Remove configure in bwd kernel launch преди 1 година
  Grigory Sizov af01244ddd Add split-kv and M<->H swap to varlen forward decoding attention (#754) преди 1 година