Struct CudaProviderBuilder 

pub struct CudaProviderBuilder { /* private fields */ }

Builds the complete production CUDA provider ownership graph.

Every provider created here owns one stream pool, one asynchronous resource stack, one runtime, and one memory manager. Providers on the same device share the canonical loaded device handle so CUDA module images are owned once per overlapping device lifetime. The same handle is shared through each provider graph; callers cannot inject mismatched allocator or device components.

Implementations

impl CudaProviderBuilder

pub fn new(device_ordinal: usize, memory_budget: MemoryBudget) -> Self

Select a CUDA device and the byte budget enforced by both the memory manager and the runtime resource stack.

pub fn with_logging_sink(self, sink: Arc<dyn LoggingSink>) -> Self

Record resource lifecycle events in sink without changing allocation or budget semantics.

pub fn with_runtime_budget_limit(self, bytes: u64) -> Self

Set the global runtime-resource byte limit independently of the memory manager limit. By default both layers use memory_budget.device_bytes.

pub fn with_stream_capacity(self, stream_capacity: usize) -> Self

Set the maximum number of pooled streams for workloads that need more concurrent lanes than the production default.

pub fn build(self) -> Result<CudaKernelProvider>

Construct and validate the complete provider-owned runtime.