Trait VariableTensorCollective 

pub trait VariableTensorCollective<B>: IntegerTensorCollective<B>
where
    B: Backend,
{
    // Required methods
    fn all_to_all_v_float(
        &self,
        value: <B as BackendTypes>::FloatTensorPrimitive,
        send_counts: &[usize],
    ) -> Result<VariableTensorExchange<<B as BackendTypes>::FloatTensorPrimitive>, Self::Error>;
    fn all_to_all_v_int(
        &self,
        value: <B as BackendTypes>::IntTensorPrimitive,
        send_counts: &[usize],
    ) -> Result<VariableTensorExchange<<B as BackendTypes>::IntTensorPrimitive>, Self::Error>;
}

Variable row exchange for routed/sharded graphs through an explicit original communicator. Every rank enters even with zero rows. Input blocks are in destination-rank order; received blocks are in source-rank order. Trailing axes must be nonzero and match across peers; no implicit expert-to-rank assignment is defined here.

Required Methods

fn all_to_all_v_float( &self, value: <B as BackendTypes>::FloatTensorPrimitive, send_counts: &[usize], ) -> Result<VariableTensorExchange<<B as BackendTypes>::FloatTensorPrimitive>, Self::Error>

Exchange actual floating row blocks, retaining original floating storage.

fn all_to_all_v_int( &self, value: <B as BackendTypes>::IntTensorPrimitive, send_counts: &[usize], ) -> Result<VariableTensorExchange<<B as BackendTypes>::IntTensorPrimitive>, Self::Error>

Exchange original U8/U32/I32/I64 row blocks without a floating surrogate.

Dyn Compatibility

This trait is not dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors