pub fn dispatch_moe<B: MoeDispatchOps>(
input: Tensor<B, 2>,
logits: Tensor<B, 2>,
bias: Option<Tensor<B, 1>>,
options: MoeOptions,
) -> Result<NativeMoeDispatched<B>, B::MoeError>Available on crate feature
api only.Expand description
Run actual original device routing and COPY dispatch before expert-parallel transport.