Commit Verlauf

Autor SHA1 Nachricht Datum
  Tri Dao 45c48afb2b Add option for WG1 to use RS MMA but WG2 using SS MMA vor 1 Tag
  Tri Dao 5458c78e6d Remove sink token vor 3 Tagen
  Tri Dao 6752d62aa4 Add dynamic splits vor 1 Woche
  Ted Zadouri 06e34f62d1 Enable MLA flag in FA3 (rope=64, latent=512) (#1504) vor 1 Woche
  Tri Dao b36ad4ef76 Use split for super long sequences that don't fit into L2 vor 2 Wochen
  Tri Dao 5378bc3204 Tile fwd_combine kernel along headdim, don't need kBlockM > 128 vor 3 Wochen
  Tri Dao 5fc5ebf82b Fix num_splits heuristic again when PackGQA vor 3 Wochen
  Tri Dao 5fab938555 Fix num_splits heuristic being called before get_pack_gqa vor 3 Wochen
  Tri Dao 4e8496a781 Support Qv vor 3 Wochen
  Tri Dao ed53b5fc4c Fix compilation for FP8 when kHeadDimV != kHeadDim vor 3 Wochen
  Tri Dao 2a204125ae Implement HeadDim_V != HeadDim_QK, support hdimQK=192, hdimV=128 vor 1 Monat
  Lu Fang 74aed78373 Replace c10::optional with std::optional in flash_attn vor 1 Monat
  Tri Dao 2ac6c986be Fix Sm80 tile_count_semaphore, adjust test tolerance vor 1 Monat
  Tri Dao e94f7e89dc Always enable PackGQA is Split to reduce compilation and binary size vor 1 Monat
  Tri Dao 40fa35acd8 Always enable PackGQA if PagedKV to reduce compilation and bin size vor 1 Monat
  Tri Dao a84a237d2a Split bwd softcap compilation units for Sm80 vor 1 Monat
  Tri Dao 1e3208566a Tune tile sizes for compilation vor 1 Monat
  Tri Dao 84f1287e42 Rename bool_constant<true> to true_type, same w bool_constant<false> vor 1 Monat
  Tri Dao df5fe55264 Change tile sizes for Sm8x to reduce stack frame vor 1 Monat
  Tri Dao 8dd0b479d5 Always enable PackGQA for Sm8x to reduce compilation and binary size vor 1 Monat
  Tri Dao 7bc3f031a4 Compile for both Sm80 and Sm90 vor 1 Monat
  Tri Dao 7a802796e1 Big refactor and update vor 1 Monat
  sclarkson 1feb711f46 Fix compilation with clang on ARM64 (#1285) vor 3 Monaten
  Kai Londenberg b443207c1f Paged Attention support for FA3 (#1268) vor 3 Monaten
  jayhshah a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236) vor 4 Monaten
  Ying Zhang dff976a84a fixes vor 6 Monaten
  Ying Zhang 7b4e68e04f hopper local attention vor 6 Monaten
  Ying Zhang db80387343 Add seqused_q in fwd / bwd and seqused_k in bwd. vor 6 Monaten
  jayhshah c92ca63268 FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173) vor 6 Monaten
  Tri Dao 5d5bfbb619 Remove contiguous checks vor 7 Monaten