Struct AllReduceRequest 

pub struct AllReduceRequest<T: NcclReduceSupported> {
    pub tensor: GpuRef<T>,
    pub op: ReduceOp,
    pub reply: Sender<Result<(), GpuError>>,
}

In-place all-reduce on a single tensor on this actor’s device.

The world actor coordinates group_start/group_end across ranks via separate BeginGroup / EndGroup messages.

Fields

tensor: GpuRef<T>

op: ReduceOp

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

Trait Implementations

impl<T: NcclReduceSupported> CollectiveDispatch for AllReduceRequest<T>

fn dtype_kind(&self) -> DispatchDType

fn device_id(&self) -> Option<u32>

fn dispatch(self: Box<Self>, ctx: &CollectiveDispatchCtx<'_>)