Struct ReduceScatterRequest 

pub struct ReduceScatterRequest<T: NcclReduceSupported> {
    pub send: GpuRef<T>,
    pub recv: GpuRef<T>,
    pub op: ReduceOp,
    pub reply: Sender<Result<(), GpuError>>,
}

ReduceScatter: each rank contributes send (length N * world_size) and receives a reduced shard of length N into recv.

Fields

send: GpuRef<T>

recv: GpuRef<T>

op: ReduceOp

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

Trait Implementations

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

fn dtype_kind(&self) -> DispatchDType

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

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