Expand description
XlogDeviceRuntime hosts one provider-owned CUDA device, stream
pool, and decorated memory-resource stack.
The canonical provider builder constructs the complete ownership graph and shares its exact handles with the memory manager. Allocation state remains provider-owned. The common admission registry compares pending storage ranges across runtimes without owning their allocator state.
Structsยง
- Conditional
Graph Stats - Execution counters for the device-controlled conditional-graph route.
- Event
Lifecycle Stats - Lifetime counters for CUDA events owned by resident graph launches.
- Resident
Completion Event - Completion event whose accounting is tied to a real cudarc event handle.
- Resident
Graph Handle Lifecycle Stats - Lifetime counters for resident CUDA graph and executable handles.
- Xlog
Device Runtime - Provider-owned CUDA device runtime.