1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
//! VoxCPM2 as a [`TtsEngine`](crate::tts::TtsEngine): the clone pipeline
//! behind one `synthesize`, and the same pipeline emitting audio chunk by
//! chunk behind `synthesize_stream`.
//!
//! A voice is a reference recording. The engine encodes every recording in a
//! voices directory once at load, so a request pays only its own prefill,
//! generation and decode. A caller encodes a recording it holds with
//! `encode_reference` and names the returned [`Reference`] in any number of
//! [`SynthRequest`]s. A Hi-Fi request also continues from a [`Prompt`]: a
//! clip encoded with `encode_prompt`, named with its transcript. Text
//! reaches the model as raw tokens, so a Malay and English code-switched
//! sentence needs no language switch.
//!
//! Every [`SynthRequest`] carries its own seed, guidance scale and solver
//! steps, or takes the load-time value of each. A buffered request retries
//! a take that runs on past its words ([`Badcase`]), with fresh noise from
//! its own seed. Requests with different
//! guidance or steps never share a generation: each waits for one of its
//! own.
//!
//! Concurrent requests share one generation. Both entry points stay `&self`
//! and blocking: they resolve their request, queue it and wait, while the
//! engine's own thread runs whatever is queued as ONE batched generation. A
//! row's noise draw is independent of the batch width and its neighbours, so
//! a batched render is bit for bit the render that row has alone.
//!
//! `types` holds the engine struct, the shared state and the constants,
//! `load` the loader and voice-directory handling, `request` the request,
//! reference and prompt types, `encode` the reference and prompt encodes,
//! `plan` resolving and tokenizing a request, `render` the caller's half
//! of a request, `batch` the queued messages and the owner thread, `queue`
//! taking work off the queue, `reply` where a row's output goes, `run` the
//! batched generation itself, and `continuous` the persistent batch behind
//! [`VoxCpm2SynthOptions::continuous_batching`](crate::voxcpm::VoxCpm2SynthOptions::continuous_batching).
pub use ;
pub use ;