baracuda-nccl-sys 0.0.1-alpha.69

Raw FFI bindings and dynamic loader for NVIDIA NCCL (multi-GPU collective communication).
Documentation

baracuda-nccl-sys

Raw FFI bindings + dynamic loader for NVIDIA NCCL — multi-GPU / multi-node collective communication (all-reduce, broadcast, all-gather, reduce-scatter, send/recv).

Symbols resolve lazily via libloading; no link-time dependency on libnccl.so.

Most users want baracuda-nccl — that crate exposes typed Communicator handles, scalar-typed collectives, p2p send/recv, group APIs, and stream-async dispatch.

What's exposed

  • ncclComm_t create / init / destroy + ncclCommSplit / ncclCommAbort / ncclCommFinalize.
  • All collectives: ncclAllReduce, ncclReduce, ncclReduceScatter, ncclAllGather, ncclBroadcast.
  • Point-to-point: ncclSend, ncclRecv.
  • Group API: ncclGroupStart, ncclGroupEnd.
  • Buffer registration: ncclMemAlloc, ncclMemFree, ncclCommRegister, ncclCommDeregister.

Platform support

NCCL is Linux-only. The loader fails fast with a clear error on Windows / macOS.

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.