voxora-backend

Cross-cutting hardware backend selection for the voxora ASR stack.

Centralises Cpu / Cuda / Metal / Vulkan selection so individual engine crates (voxora-whisper, voxora-qwen3asr, future parakeet / voxtral adapters, …) no longer each declare their own cpu / cuda / metal / vulkan feature wiring.

ASR-specific: only models hardware backends that exist in the voxora speech-recognition stack. Does not abstract LLama.cpp or any other non-ASR backend.

Phase 0 surface

• [BackendKind] — re-exported from voxora-engine.
• [Capabilities] — compile-time + runtime snapshot.
• [best_device] / [best_device_or_error] — current "always returns Cpu" implementation.

candle Cargo feature

voxora-backend 0.3.0 adds an opt-in candle Cargo feature. When enabled, [best_device] and [detect] use candle-core's runtime probe (CUDA → Metal → CPU) instead of always returning Cpu.

voxora-backend = { version = "0.3", features = ["candle"] }

The feature is off by default to avoid pulling in candle-core's transitive deps (cudarc, objc2-metal, …) for users who don't need runtime detection. It will be promoted to default-on in 0.3.1 once candle integration is validated on more platforms.

CI caveat: the GitHub Actions matrix runs with default features only. The candle feature is not exercised by CI; if you depend on it, please verify locally before bumping voxora-backend.

Without the candle feature the behaviour is unchanged: both [best_device] and [detect] return Cpu as the safe fallback.

Cross-engine hardware support

The cross-engine hardware / GPU support matrix (which engine ships which backend, CUDA compute capability per backend, the voxora-bridge forwarding rules, and the docs.rs sandbox caveat) is documented in docs/GPU_SUPPORT.md. This crate is the runtime picker that GPU-support doc defers to when consumers want "use whatever's available" instead of pinning a device at compile time.