Expand description
Cross-cutting hardware backend selection for voxora ASR engines.
Centralises Cpu / Cuda / Metal / Vulkan selection so engines no
longer each declare their own cpu / cuda / metal / vulkan
feature wiring.
ASR-specific: this crate only models hardware backends that are relevant to the voxora speech-recognition stack. It does not abstract LLama.cpp or any other non-ASR backend.
§Phase 0 surface
Today the crate only exposes:
BackendKind(re-exported fromvoxora-engine).Capabilities— compile-time + runtime snapshot.best_device/best_device_or_error— current “always returns CPU” implementation.
The real candle-backed detection lands in 0.2.x once voxora-backend
gains a candle feature flag.
Re-exports§
pub use capabilities::Capabilities;pub use device::best_device;pub use device::best_device_or_error;pub use device::detect;pub use error::BackendError;
Modules§
- capabilities
- Snapshot of what backends the current binary supports.
- device
- Best-effort device selection.
- error
- Errors returned by voxora-backend.
- kind
- Hardware backend kind.
Enums§
- Backend
Kind - Compile-time hardware backend the engine was built against.