Expand description
GpuRef<T> — opaque, message-friendly handle to a GPU buffer (§5.8).
Send + Sync + 'static with no lifetime parameters, so it composes
freely in any actor message type. Validity is checked at runtime by
comparing a generation token against DeviceState.generation, which
is bumped whenever the underlying CudaContext is rebuilt (§5.11
supervision).
Cross-node serialisation (GpuToken { node_id, device_id, buffer_id, generation } — §5.5) is intentionally not implemented in F1; it
lands with the F4 cluster/NCCL story. F1 GpuRef is local-only.
Structs§
- GpuRef
- A live device-buffer handle.