Struct Model 

pub struct Model {
    pub language_model: TextModel,
    pub vision_tower: VisionTower,
    pub embed_vision: MultimodalEmbedder,
    pub audio_tower: Option<AudioModel>,
    pub embed_audio: Option<MultimodalEmbedder>,
    pub cfg: Gemma4Config,
}

Full Gemma4 multimodal model.

Fields

language_model: TextModel

vision_tower: VisionTower

embed_vision: MultimodalEmbedder

audio_tower: Option<AudioModel>

embed_audio: Option<MultimodalEmbedder>

cfg: Gemma4Config

Implementations

impl Model

pub fn new(cfg: &Gemma4Config, vb: VarBuilder<'_>) -> Result<Self>

pub fn forward( &mut self, input_ids: &Tensor, seqlen_offset: usize, ) -> Result<Tensor>

Text-only forward pass.

pub fn forward_multimodal( &mut self, input_ids: &Tensor, pixel_values: Option<&[Tensor]>, audio_mel: Option<&Tensor>, audio_mel_mask: Option<&Tensor>, seqlen_offset: usize, ) -> Result<Tensor>

Forward with multimodal inputs.

pixel_values: optional batch of images, each (1, C, H, W). audio_mel: optional (batch, time, mel_bins) mel spectrogram. audio_mel_mask: optional (batch, time) mask (1.0 = padding).

pub fn clear_kv_cache(&mut self)