Expand description
2026-09-26: Collective-communication backends behind one trait,
CommBackend: SingleGpuBackend, where every operation is a no-op,
and, with the nccl feature, NcclBackend. No other crate calls NCCL.
Owner: metrale-comm. Invariants: none beyond the types.
Re-exports§
pub use nccl_backend::NcclBackend;
Modules§
- nccl_
backend - 2026-09-26:
NcclBackend, the multi-rankCommBackendover NCCL. Bootstrap: rank 0 creates an NCCL unique id and sends it over TCP to every other rank, then all ranks callncclCommInitRank.
Structs§
- Single
GpuBackend - 2026-09-26: One rank: every collective and point-to-point operation
returns
Okwithout doing anything,rankis 0 andworld_sizeis 1.
Traits§
- Comm
Backend - 2026-09-26: Collective and point-to-point operations on raw device
pointers and byte counts. A pointer is a
u64(a CUDACUdeviceptr), so this crate does not depend on metrale-gpu-runtime.