Module multicast 

Multicast objects (CUDA 12.0+).

A multicast object is a single VMM handle bound to multiple devices; writes through the handle are implicitly replicated across them. Useful for NVLink-connected GPUs (A100/H100) on all-reduce-like workloads when you don’t want to go through NCCL.

Not supported on older drivers — returns crate::Error::FeatureNotSupported.

Structs

MulticastObject

A multicast object. Drop releases it via cudaMemRelease.

MulticastProp

Properties passed to MulticastObject::new. Layout matches cudaMulticastObjectProp.

Functions

multicast_granularity

Report the granularity (min alignment / min size) for multicast objects with the given props. option: 0 = minimum, 1 = recommended.