Commit History

作者 SHA1 備註 提交日期
  Piotr Wilkin 5306640300 All's well that ends in a well 3 月之前
  Piotr Wilkin 232ec56251 Yes, I finally managed to implement it with ssm_conv :> 3 月之前
  Piotr Wilkin aa8d6a21a3 Remove extra files cont. 3 月之前
  Piotr Wilkin e9a98f2af9 Remove extra files 3 月之前
  Piotr Wilkin 22ee5a971b Add gate_sigmoid to callback 3 月之前
  Piotr Wilkin ce87b7d78e Yup, it's NeoX 3 月之前
  Piotr Wilkin df0b5bcf30 Proper order of attention operations 3 月之前
  Piotr Wilkin 54712b8664 Oh, forgot to commit 3 月之前
  Piotr Wilkin 17240eafc0 Order stuff around 3 月之前
  Piotr Wilkin 1579bcb202 What am I missing? :/ 3 月之前
  Piotr Wilkin 0a9244acd0 The optimization worked even too well ;) 4 月之前
  Piotr Wilkin 8ddaf251ae Fix some state regressions... still wip 4 月之前
  Piotr Wilkin 6942c85cf8 Oh, actually set n_tasks as well :P 4 月之前
  Piotr Wilkin 477c1616ad Parallelize delta_net 4 月之前
  Piotr Wilkin 4ef6f337de Proper multi-sequence convolution calculation, corrected (?) state management 4 月之前
  Piotr Wilkin 5f5e30007c Dilution n_seqs -> 1 4 月之前
  Piotr Wilkin eb0a15fc9b n_tokens -> n_seq_tokens 4 月之前
  Piotr Wilkin ee52fe36f3 Modify sanity check to handle hybrid models 4 月之前
  Piotr Wilkin 0dd6110fdc v1.0 4 月之前
  Piotr Wilkin adcbd9428f Linear layer output convergence 4 月之前
  Piotr Wilkin 666fc0583d Parity on delta! 4 月之前
  Piotr Wilkin a2c7b6794e Proper handling for n_tokens > GGML_DELTA_NET_CHUNK 4 月之前
  Piotr Wilkin c1e46f62fa Achieve pre-chunk-attention parity; remove most of the LLM generated crap 4 月之前
  Piotr Wilkin c87e8d550c Tensor preparation for delta_net complete 4 月之前
  Piotr Wilkin 7ec2df64a4 Added: tri, cumsum. Still a mess. 4 月之前
  Piotr Wilkin 6d0ad37cf4 Fix QKV extraction post-convolution 4 月之前
  Piotr Wilkin 845a3d7166 Convolution 4 月之前
  Piotr Wilkin 638057a29b Transpose input for convolution 4 月之前
  Piotr Wilkin 835d389fc5 Fix BA views as well 4 月之前
  Piotr Wilkin 594c1f98ef QKV splits done right 4 月之前