Struct BackboneModel 

pub struct BackboneModel {
    pub seed: Option<u32>,
    /* private fields */
}

NeuTTS backbone — RLX Llama-3.2 runner over a llama-tagged GGUF.

Fields

seed: Option<u32>

Implementations

impl BackboneModel

pub fn load(path: &Path, n_ctx: u32) -> Result<BackboneModel, Error>

pub fn load_on( path: &Path, n_ctx: u32, device: Device, ) -> Result<BackboneModel, Error>

Load the GGUF backbone on a specific execution device.

pub fn load_greedy_parity( path: &Path, n_ctx: u32, ) -> Result<BackboneModel, Error>

F32 dequant + incremental greedy (tail parity vs llama-cpp Q4).

pub fn load_greedy_parity_on( path: &Path, n_ctx: u32, device: Device, ) -> Result<BackboneModel, Error>

Greedy parity load on a specific execution device.

pub fn generate( &self, prompt: &str, max_new_tokens: u32, ) -> Result<String, Error>

pub fn generate_streaming<F>( &self, prompt: &str, max_new_tokens: u32, on_piece: F, ) -> Result<(), Error>

where F: FnMut(&str) -> Result<(), Error>,

pub fn generate_greedy_ids( &self, prompt: &str, max_new_tokens: u32, ) -> Result<Vec<u32>, Error>

Greedy token IDs for parity tests (same GGUF vocab as production).

pub fn generate_greedy_ids_from_prompt( &self, prompt_ids: &[u32], max_new_tokens: u32, ) -> Result<Vec<u32>, Error>

Greedy continuation for parity tests.

[load_greedy_parity] uses KV-cached Llama32Runner::generate (F32 weights, MSVC uses oneshot decode in step_cached). Production [load] uses packed Q4. Debug: NEUTTS_GREEDY_INCREMENTAL=1 or NEUTTS_GREEDY_PREDICT_LOGITS=1.

pub fn generate_greedy( &self, prompt: &str, max_new_tokens: u32, ) -> Result<String, Error>

Greedy text generation for parity tests (deterministic vs llama.cpp reference).