Skip to content

generic prompting in interface - #178

Open
cjpais wants to merge 32 commits into
mainfrom
generic-prompting
Open

cjpais wants to merge 32 commits into
mainfrom
generic-prompting

Conversation

@cjpais

@cjpais cjpais commented Sep 27, 2026

Copy link
Copy Markdown
Contributor

No description provided.

…ocabulary, prefix or diarization is requested
…unless vocabulary, prefix or diarization is requested"

This reverts commit 7ccfac6.
- typescript: marshal vocabulary as a char* array (it reached C as one empty
  string: 2+ terms were rejected, 1 was ignored); validate terms; free the
  buffer after each call
- python: accept iterators for vocabulary, reject sets/dicts (priority order)
- all: reject NUL characters in string options instead of truncating
  (Rust already errored)
- rust: #[non_exhaustive] on Task and Feature
- model-gated prompting tests for typescript, rust and swift; python tests
  tightened (unsupported prefix on the streaming canary, batch/stream cases)
- terse prompting docs in each binding README and doc comments
…whisper: full prompt on later ALL_SEGMENTS windows

- causal_lm: kv_grow reallocates a single-utterance cache and copies the
  filled rows, for decodes whose length is not known up front
- voxtral: free-text (INSTRUCT) output is no longer capped by the
  audio-derived transcript budget (448 tokens on short clips); it runs until
  EOS, a repetition stop or the context ceiling (n_ctx still caps it). The
  cache starts at the transcript size and grows as needed. INSTRUCT batches
  run serially, since the packed batched cache cannot grow per row
- whisper: with a transcript prefix, only the first window's prompt is
  shortened to make room for it; later ALL_SEGMENTS windows get the full slot

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant