Enum CollectiveMsg 

pub enum CollectiveMsg {
    Op(Box<dyn CollectiveDispatch>),
    BeginGroup {
        reply: Sender<Result<(), GpuError>>,
    },
    EndGroup {
        reply: Sender<Result<(), GpuError>>,
    },
    QueryCapabilities {
        reply: Sender<NcclCapabilities>,
    },
    AllReduceF32 {
        tensor: GpuRef<f32>,
        op: ReduceOp,
        reply: Sender<Result<(), GpuError>>,
    },
    BroadcastF32 {
        data: GpuRef<f32>,
        root: usize,
        reply: Sender<Result<(), GpuError>>,
    },
}

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.

Variants

Op(Box<dyn CollectiveDispatch>)

Boxed dispatch: any typed *Request<T: NcclReduceSupported> implements CollectiveDispatch and ships through this variant. New ops (AllGather, ReduceScatter, AllToAll, …) all arrive this way.

BeginGroup

Begin a group call. Issues ncclGroupStart on this rank.

Fields

reply: Sender<Result<(), GpuError>>

EndGroup

End a group call. Issues ncclGroupEnd on this rank.

Fields

reply: Sender<Result<(), GpuError>>

QueryCapabilities

Probe the loaded NCCL library for capabilities (version, fp8 / NVLS / SHARP support). Returns zeros if NCCL isn’t initialised on this host.

Fields

reply: Sender<NcclCapabilities>

AllReduceF32

👎Deprecated: use CollectiveMsg::Op(Box::new(AllReduceRequest::<f32> { … })) instead

Legacy alias preserved for back-compat. New callers should build AllReduceRequest<f32> and ship via CollectiveMsg::Op.

Fields

tensor: GpuRef<f32>

👎Deprecated: use CollectiveMsg::Op(Box::new(AllReduceRequest::<f32> { … })) instead

op: ReduceOp

👎Deprecated: use CollectiveMsg::Op(Box::new(AllReduceRequest::<f32> { … })) instead

reply: Sender<Result<(), GpuError>>

👎Deprecated: use CollectiveMsg::Op(Box::new(AllReduceRequest::<f32> { … })) instead

BroadcastF32

👎Deprecated: use CollectiveMsg::Op(Box::new(BroadcastRequest::<f32> { … })) instead

Legacy alias preserved for back-compat. New callers should build BroadcastRequest<f32> and ship via CollectiveMsg::Op.

Fields

data: GpuRef<f32>

👎Deprecated: use CollectiveMsg::Op(Box::new(BroadcastRequest::<f32> { … })) instead

root: usize

👎Deprecated: use CollectiveMsg::Op(Box::new(BroadcastRequest::<f32> { … })) instead

reply: Sender<Result<(), GpuError>>

👎Deprecated: use CollectiveMsg::Op(Box::new(BroadcastRequest::<f32> { … })) instead