Crate baracuda_nccl 

Safe Rust wrappers for NVIDIA NCCL (multi-GPU collective communication).

Layered on top of baracuda-nccl-sys. Use this crate directly for typed, RAII-managed communicators + collectives; reach for -sys only when adding a function the safe layer doesn’t expose yet.

Scope

• Communicator lifecycle: single-process multi-GPU via ncclCommInitAll, multi-process via ncclCommInitRank + UniqueId exchange, communicator destruction, error querying.
• All collectives: all_reduce, all_gather, broadcast, reduce, reduce_scatter, send / recv (point-to-point used by baracuda-kernels’s Ring Attention K/V chunk rotation).
• Group operations: group_start / group_end for batched collective launches (essential for Megatron-LM TP all-reduce patterns).
• Communicator features: stream binding, error checking, abort + finalize for graceful shutdown.
• Datatype helpers: f32/f64/f16/bf16/u8/i32/i64/u64 reduction support via DataType enum.
• Reduction ops: sum, prod, min, max, avg, pre_mul_sum.

Platform support

NCCL is primarily a Linux library. Windows has experimental support in newer NCCL versions but is uncommon. The crate compiles on Windows but Communicator::init_all returns LoaderError::LibraryNotFound at runtime on hosts without NCCL — single-device callers can detect this and fall back gracefully.

Usage with baracuda-kernels

baracuda-kernels’s Ring Attention plan (Phase 56, ring_attention feature) and Megatron-LM TP primitives (Phase 57, megatron_tp feature) both consume this crate. Direct callers commonly use the communicator for synchronous data-parallel training all-reduce.

Re-exports

pub use RedOp as NcclReduceOp;

pub use UniqueId as NcclUniqueId;

pub use NcclScalar as NcclDataType;

Structs

Communicator

A NCCL communicator — one rank’s view of a distributed group.

NcclMem

NCCL-managed device allocation. Drop calls ncclMemFree.

UniqueId

A 128-byte opaque identifier for establishing a multi-process NCCL communicator. One process calls UniqueId::new and distributes the bytes to all other processes via a user-provided channel (TCP, MPI, …); every process then calls Communicator::init_rank with the same id.

Enums

RawNcclDataType

Re-export the raw ncclDataType_t enum so callers can pattern-match or pass it to lower-level helpers if needed. NCCL element data type.

RedOp

Reduction operation for all_reduce / reduce.

ScalarResidence

Where the scalar passed to Communicator::create_pre_mul_sum lives.

Traits

NcclScalar

Element type for NCCL buffers. Implemented by baracuda-types primitives via a sealed trait.

Functions

all_reduce

All-reduce: each rank sends send and receives the per-element reduction (across every rank) into recv. In-place use (send == recv) is legal.

broadcast

Broadcast the data at root’s send buffer to every rank’s recv buffer.

error_string

Human-readable name for a status code.

group_end

End the current collective group.

group_start

Begin a group of collectives that must be submitted atomically (e.g. in single-process multi-GPU all-reduce).

version

NCCL library version as a packed integer (e.g. 22100 for NCCL 2.21.0).

Type Aliases

Error

Error type for NCCL operations.

Result

Result alias.