The TensorPipe project provides a tensor-aware channel to transfer rich objects from one process to another while using the fastest transport for the tensors contained therein (e.g., CUDA device-to-device copy).