Sparse All-Reduce in PyTorch | Speechmatics