Struct CachedBackend 

pub struct CachedBackend<T: 'static> { /* private fields */ }

Probe the CUDA backend for label and run a backend-specific build step on the resolved handles.

This is probe_cuda_backend plus the one piece that genuinely differs between backends: the NVRTC compile (and any per-backend cache construction). The runtime resolution, context creation, and stream selection — together with their uniform, label-attributed error messages — live in the shared probe; build receives the resolved CudaBackendParts (so it can clone the Arc<CudaContext> / Arc<CudaStream> it needs) and returns the backend’s own state T. A process-wide backend for one kernel family: the context and stream from probe_cuda_backend plus whatever the family builds on top of them (compiled modules, device limits), probed and built once and then shared by every caller for the life of the process. A failed probe is cached too, so a host without a usable device answers every later call with the same refusal instead of re-probing.

Declare one per family as a static and read it through CachedBackend::get_or_probe; five kernel families used to spell this OnceLock<Result<_, GpuError>> protocol out by hand (#2470).

Implementations

impl<T: 'static> CachedBackend<T>

pub const fn new() -> Self

pub fn get_or_probe<F>( &'static self, label: &'static str, build: F, ) -> Result<&'static T, GpuError>

where F: FnOnce(CudaBackendParts) -> Result<T, GpuError>,

The shared backend, probing and building it on first use.

Trait Implementations

impl<T: 'static> Default for CachedBackend<T>

fn default() -> Self

Returns the “default value” for a type.