Module cuda 

GPU-facing configuration and transfer descriptors.

The crate keeps these types in the public API so storage callers can describe KV-cache chunk transfers without depending on a CUDA runtime in the core crate. The actual GPU execution layer is intentionally outside this package.

Structs

CudaChunkTransferDescriptor

Precomputed routing metadata for a chunk that should be transferred to a GPU destination in layer order.

CudaChunkTransferHit

CudaConfig

Runtime configuration for the optional CUDA/GPU tier.

CudaSessionTransferRequest

A session-scoped transfer request for streaming KV chunks in layer order to a GPU-facing consumer.

CudaSessionTransferStats

Aggregate outcome for a session-scoped streaming transfer.

Enums

CudaSessionChunkEvent