turbospark-runtime 0.3.0

The raw-completion prefill+decode loop: wires a LogitProducer, the tokenizer, and the selection sampler into one token generation loop.
1
2
3
4
5
6
7
8
9
10
11

//! Model family execution modules for real and synthetic checkpoints.

pub(crate) mod deepseek2;
pub(crate) mod gemma4;
pub(crate) mod gptoss;
pub(crate) mod llama;
pub(crate) mod museglimmer;
pub(crate) mod qwen;
pub(crate) mod qwen4;
pub(crate) mod spark;
pub(crate) mod synthetic;