Expand description
Async collective operations: AllReduce via dedicated comm threads.
Structs§
- AllReduce
Handle - Non-blocking AllReduce handle.
- Collective
Barrier - Cross-rank barrier for summing f32 vectors.
- Comm
Request - Request from the compute thread to the comm thread.
- Comm
Thread - Dedicated communication thread for async AllReduce.
Functions§
- async_
allreduce - Submit a non-blocking AllReduce via the comm thread.