//! # Driver support: the model-generic parts of a generation driver
//!
//! What a driver that streams a language model's output needs, whatever the
//! model: none of it knows a tokenizer, a template or a weight.
//!
//! - [`TextStream`]: a reply's text as it grows, one delta per token, never
//! splitting a character whose bytes span two tokens; over any
//! [`Detokenizer`](crate::kits::support::tokens::Detokenizer).
//! - [`StopStrings`]: stop strings matched across those deltas, holding back
//! what might yet become one.
//! - [`CachedHistory`]: the tokens a
//! [`KVCache`](crate::ops::transformers::attention::KVCache) holds, and the
//! rewind to the prefix the next prompt shares with them
//! ([`KVCache::truncate`](crate::ops::transformers::attention::KVCache::truncate)),
//! so a multi-turn session prefills only what changed.
//! - [`CachedLm`]: a language model a driver decodes by token ids, whatever its
//! forward: its decode state, its prefill, its rewind to a [`CachedHistory`].
//!
//! The searches a driver runs are
//! [`kits::support::decoding`](crate::kits::support::decoding)'s.
pub use *;
pub use *;
pub use *;
pub use *;