Module collective 

CollectiveActor — wraps an [cudarc::nccl::Comm] for one rank within an NcclWorldActor group.

Phase 2 NCCL slice: full collective surface (AllReduce, AllGather, ReduceScatter, AllToAll(v), Reduce, Broadcast), point-to-point Send/Recv, typed group scope guard, NVLS/SHARP/fp8 capability probe, and a custom PreMulSum reduce op. dtype-generic via the NcclReduceSupported marker (defined here until Phase 0 lands).

Each CollectiveActor is bound to one specific crate::device::DeviceState (one rank in the NCCL world). The parent NcclWorldActor spawns N of these (one per device) and routes messages to all of them in a group_start/group_end pair where appropriate.

Re-exports

pub use all_to_all::AllToAllRequest;

pub use all_to_all::AllToAllvRequest;

pub use allgather::AllGatherRequest;

pub use allreduce::AllReduceRequest;

pub use broadcast::BroadcastRequest;

pub use capabilities::probe_capabilities;

pub use capabilities::NcclCapabilities;

pub use custom_op::PreMulSumOp;

pub use group::GroupGuard;

pub use p2p::RecvRequest;

pub use p2p::SendRequest;

pub use reduce::ReduceRequest;

pub use reduce_scatter::ReduceScatterRequest;

Modules

all_to_all

Typed AllToAll / AllToAllv requests.

allgather

Typed AllGather request — generic over T: NcclReduceSupported.

allreduce

Typed AllReduce request. Generic over T: NcclReduceSupported (any of f32/f64/i8/u8/i32/u32/i64/u64; f16/bf16 with f16).

broadcast

Typed Broadcast request — generic over T: NcclReduceSupported.

capabilities

Runtime probe for NCCL capabilities.

custom_op

Custom reduce ops — currently ncclRedOpCreatePreMulSum.

group

Typed scope guard around ncclGroupStart / ncclGroupEnd.

p2p

Typed point-to-point Send / Recv requests.

reduce

Typed Reduce request: reduce-to-root variant of AllReduce.

reduce_scatter

Typed ReduceScatter request — generic over T: NcclReduceSupported.

Structs

CollectiveActor

Enums

CollectiveMsg

Public message surface for the CollectiveActor. Hot path goes through CollectiveMsg::Op which carries a boxed CollectiveDispatch; the legacy AllReduceF32 / BroadcastF32 variants remain for back-compat and route through the same machinery.

ReduceOp

Traits

NcclReduceSupported

Marker for dtypes carried in NCCL collectives. Mirrors the NcclReduceSupported marker that the Phase 0 dtype.rs will host — defined locally here so the NCCL slice can ship before Phase 0 fully lands. The set matches NCCL’s reduce-supported types: f32, f64, f16, bf16, i8, u8, i32, u32, i64, u64. fp8 e4m3/e5m2 are behind nccl-fp8 and require NCCL >= 2.20.