Expand description
CUDA device management
This module keeps XLOG’s historical single-stream device abstraction while targeting cudarc’s newer CUDA 13-capable context/stream APIs.
Loading/replacing modules and releasing their last owners can wait for execution-event completion, directly or through deferred capture cleanup. Callers must satisfy the cold-lifecycle progress contract.
Structs§
- Cuda
Device - CUDA device wrapper for GPU operations.
- Cuda
Device Inner - Cuda
Function - Kernel handle bound to XLOG’s default CUDA stream. Retains the exact loaded module independently of its registry name. Releasing the last module owner follows the potentially blocking cold-lifecycle progress contract.