rlx-text
Public surface for downstream LM apps:
- Chat templates — Jinja-rendered chat templates loaded inline or from
GGUF metadata (
tokenizer.chat_template,tokenizer.ggml.bos_token_id, …). - Tokenizer wrapper — thin façade over the
tokenizerscrate so model crates don't all rebuild the same encode/decode glue. - Sampling — re-exports
rlx_runtime::SampleOptsplus helpers (argmax, top-k, top-p, repetition penalty).
Replaces the chat/sampling/tokenizer pieces that used to live in
rlx-models/rlx-cli, breaking the "runners depend on the CLI helper crate"
coupling.
License
GPL-3.0-only.