Expand description
Incin’s execution backends: the built-in CPU backend plus feature-gated CUDA, WGPU, Metal, and external Candle implementations behind one dispatch surface, with the capability registries and codegen helpers backend authors share.
Re-exports§
pub use dispatch::DispatchBackend;pub use detect::detect_device;pub use detect::detect_device_in;pub use simd::simd_lanes;pub use telemetry::set_emitter;
Modules§
- backend_
kind - Device-to-backend selection for the unified
IncinBackendfacade. - capability
- Authoritative native-backend capability registrations.
- capability_
docs - The generated support tables in
docs/capabilities.md(UX-013). The support tables indocs/capabilities.md, rendered from the registrations incrate::capability. - codegen
- Unified cross-backend kernel emitter module (PRF-007).
- cpu
- Incin Cpu
- detect
- Runtime detection of the best device this machine can actually run on. Runtime detection of the best device this machine can actually run on.
- dispatch
- Runtime backend selection used by
IncinBackend<_, Dyn>. - dist
- Collective transport contracts and optional implementations. Distributed collective transport contracts.
- external
- Third-party backend integrations, and the conformance suite an author of one runs against their own backend.
- nn_
target - Central target-aware parameter materialization and NN layer initialization on
TensorTarget. - prelude
- Single import surface for backend authors.
- simd
- Compile-time SIMD lane-width resolution for type-specialized kernels.
- target
- A device value as the user-facing allocation target.
- telemetry
- Optional telemetry integration for
incin-backends.
Type Aliases§
- Engine
Backend - Backend parameterized by a target execution engine.
- Incin
Backend - Unified backend selected by device.
- Native
Backend - Backend bound directly to one native device kind.