Struct MimoAudio 

pub struct MimoAudio {
    pub tokenizer: AudioTokenizer,
    pub encoder: AudioEncoder,
    pub bf16_codebooks: bool,
    /* private fields */
}

The MiMo audio towers.

Fields

tokenizer: AudioTokenizer

encoder: AudioEncoder

bf16_codebooks: bool

Quantize against bf16-rounded codebooks, as the bf16 serving stacks do (default on; CMF_MIMO_AUDIO_BF16_BOOKS=0 turns it off).

Implementations

impl MimoAudio

pub fn present_in(model: &CmfModel) -> bool

Whether a CMF carries the audio towers (a companion or a single-file multimodal container).

pub fn from_model(model: &Arc<CmfModel>) -> Result<Self, String>

Load from a CMF holding the source tensor names (the <stem>.mm.cmf companion or a single-file multimodal container).

pub fn from_hf_dir(dir: &Path) -> Result<Self, String>

Development loader: read the tensors straight from the HF checkpoint directory (config.json, model.safetensors.index.json and its shards, audio_tokenizer/), exact f32 from BF16.

pub fn pool(&self) -> Option<&Pool>

pub fn wav_to_mel(&self, bytes: &[u8]) -> Result<(Vec<f32>, usize), String>

WAV bytes → 24 kHz mono → log-mel [M][128].

pub fn features(&self, mel: &[f32], m: usize) -> Result<Vec<f32>, String>

Pre-RVQ features of every segment, concatenated [codes][d].

pub fn encode_mel(&self, mel: &[f32], m: usize) -> Result<AudioCodes, String>

Log-mel → tokenizer codes, each 6000-frame segment on its own.

pub fn embed_codes(&self, codes: &AudioCodes) -> Result<AudioEmbeds, String>

Codes → LLM embedding rows.

pub fn embed_wav(&self, bytes: &[u8]) -> Result<AudioEmbeds, String>

WAV bytes → LLM embedding rows (one per <|audio_pad|>).