onde
On-device inference abstraction layer for SplitFire AI.
onde wraps mistral.rs model management so that individual Tauri apps remain thin command wrappers. It handles:
- Model metadata — download sizes, display names, organisations, descriptions.
- HuggingFace token resolution — build-time literal or on-disk cache (required on iOS where the filesystem is sandboxed).
- HuggingFace hub cache — inspection, repair, and model download with a progress-callback API decoupled from Tauri.
Platform support
| Platform | Backend |
|---|---|
| macOS | Metal-accelerated inference |
| iOS | Metal-accelerated inference |
| Android | CPU-only GGUF inference via candle |
| Windows | CPU-only baseline (CUDA opt-in via CI) |
Usage
Add to your Cargo.toml:
[]
= "0.1.0"
Re-exports are available as onde::mistralrs, onde::hf_hub, and onde::mistralrs_core so downstream crates do not need direct dependencies on those crates.
License
MIT