Commit Verlauf

Autor SHA1 Nachricht Datum
  AlpinDale a4cbcfe59f feat: disable logprob serialization to CPU for spec decode vor 5 Monaten
  AlpinDale 3a53ff1e01 fix: raise an error for no draft token case when draft_tp>1 vor 5 Monaten
  AlpinDale 2c653a2268 fix: make speculative decoding work with per-request seed vor 5 Monaten
  AlpinDale 16dff9babc chore: enable bonus token in spec decoding for KV cache based models vor 5 Monaten
  AlpinDale d9f4c36edd feat: Medusa speculative decoding support (#590) vor 5 Monaten
  AlpinDale dd378ea063 feat: MLPSpeculator with tensor parallel vor 6 Monaten
  AlpinDale 7253e9052d feat: integrate typical acceptance sampling for spec decoding vor 6 Monaten
  AlpinDale cdff8e89f9 feat: introduce `DraftModelRunner` vor 6 Monaten
  AlpinDale b6ff0623a6 chore: clean up branding vor 6 Monaten
  AlpinDale abbb730607 feat: support draft model on different tensor parallel size vor 6 Monaten
  AlpinDale af43576da0 feat: add MLPSpeculator speculative decoding support (#572) vor 6 Monaten
  AlpinDale 313e6e1ec7 feat: add typical acceptance sampling vor 6 Monaten
  AlpinDale 4d1e613804 chore: minor simplifications vor 6 Monaten
  AlpinDale e0886ee929 feat: add `ProposerWorkerBase` abstract class vor 6 Monaten
  AlpinDale 344ddaac5a properly disable speculative decoding vor 6 Monaten
  AlpinDale 5b0c11d190 support pipeline parallel pynccl groups vor 6 Monaten
  AlpinDale de62ceb18c refactor: eliminate parallel worker per-step task scheduling overhead vor 6 Monaten
  AlpinDale 236be273e5 feat: tensor parallel speculative decoding (#554) vor 6 Monaten
  AlpinDale 197a6d2c16 auto disable speculative decoding by the running queue size vor 6 Monaten
  AlpinDale 438f5bdce9 fix ngrams vor 6 Monaten
  AlpinDale ef733aee43 implement ExecuteModelData to reduce executor complexity vor 6 Monaten
  AlpinDale 79901b76de logprobs for target model (spec decoding) vor 6 Monaten
  AlpinDale 723c6acb84 re-add ngram speculative decoding vor 6 Monaten
  AlpinDale 7bcf4c3fc9 centralize gpu worker construction vor 6 Monaten
  AlpinDale 9d81716bfd [v0.5.3] Release Candidate (#388) vor 9 Monaten
  AlpinDale f8dfac6372 chore: attention refactor and upstream sync apr01 (#365) vor 10 Monaten