rullama 0.3.0

Browser-resident Gemma 4 inference: pure Rust → WebAssembly + WebGPU. Loads Ollama's on-disk GGUF blobs and runs the forward pass on the local GPU via hand-written WGSL.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

//! Multimodal towers for Gemma 4 (vision + audio).
//!
//! Mirrors Ollama's `model/models/gemma4/{model_vision,model_audio,process_image,process_audio}.go`.
//! The text language model is unchanged — multimodal features become "soft tokens"
//! injected into the residual stream via `Forward::step_with_embedding`.

pub mod audio;
pub mod audio_features;
pub mod audio_gpu;
pub mod vision;

pub use audio::{AudioConfig, AudioPrefix};
pub use audio_features::{MEL_BINS, MelEngine, SAMPLE_RATE, decode_wav};
pub use audio_gpu::GpuAudioForward;
pub use vision::{VisionConfig, VisionForward};