Commit Verlauf

Autor SHA1 Nachricht Datum
  Piotr Wilkin aa8d6a21a3 Remove extra files cont. vor 3 Monaten
  Piotr Wilkin e9a98f2af9 Remove extra files vor 3 Monaten
  Piotr Wilkin 22ee5a971b Add gate_sigmoid to callback vor 3 Monaten
  Piotr Wilkin ce87b7d78e Yup, it's NeoX vor 3 Monaten
  Piotr Wilkin df0b5bcf30 Proper order of attention operations vor 3 Monaten
  Piotr Wilkin 54712b8664 Oh, forgot to commit vor 3 Monaten
  Piotr Wilkin 17240eafc0 Order stuff around vor 3 Monaten
  Piotr Wilkin 1579bcb202 What am I missing? :/ vor 3 Monaten
  Piotr Wilkin 0a9244acd0 The optimization worked even too well ;) vor 3 Monaten
  Piotr Wilkin 8ddaf251ae Fix some state regressions... still wip vor 3 Monaten
  Piotr Wilkin 6942c85cf8 Oh, actually set n_tasks as well :P vor 3 Monaten
  Piotr Wilkin 477c1616ad Parallelize delta_net vor 3 Monaten
  Piotr Wilkin 4ef6f337de Proper multi-sequence convolution calculation, corrected (?) state management vor 3 Monaten
  Piotr Wilkin 5f5e30007c Dilution n_seqs -> 1 vor 3 Monaten
  Piotr Wilkin eb0a15fc9b n_tokens -> n_seq_tokens vor 3 Monaten
  Piotr Wilkin ee52fe36f3 Modify sanity check to handle hybrid models vor 3 Monaten
  Piotr Wilkin 0dd6110fdc v1.0 vor 3 Monaten
  Piotr Wilkin adcbd9428f Linear layer output convergence vor 3 Monaten
  Piotr Wilkin 666fc0583d Parity on delta! vor 3 Monaten
  Piotr Wilkin a2c7b6794e Proper handling for n_tokens > GGML_DELTA_NET_CHUNK vor 4 Monaten
  Piotr Wilkin c1e46f62fa Achieve pre-chunk-attention parity; remove most of the LLM generated crap vor 4 Monaten
  Piotr Wilkin c87e8d550c Tensor preparation for delta_net complete vor 4 Monaten
  Piotr Wilkin 7ec2df64a4 Added: tri, cumsum. Still a mess. vor 4 Monaten
  Piotr Wilkin 6d0ad37cf4 Fix QKV extraction post-convolution vor 4 Monaten
  Piotr Wilkin 845a3d7166 Convolution vor 4 Monaten
  Piotr Wilkin 638057a29b Transpose input for convolution vor 4 Monaten
  Piotr Wilkin 835d389fc5 Fix BA views as well vor 4 Monaten
  Piotr Wilkin 594c1f98ef QKV splits done right vor 4 Monaten
  Piotr Wilkin dbd4d97cf2 Fix cb calls vor 4 Monaten
  Piotr Wilkin 32dcee47ef Some attempts to get the convolution input right. vor 4 Monaten