Transfering CUDA tensors between candle and tch-rs without synchronization

Is there a way to transfer CUDA tensors between candle and tch-rs without requiring GPU/CPU synchronisation?

I am doing reinforcement learning and currently I have a ResNet that I am training in tch-rs. I have found that the main bottleneck of my implementation is GPU/CPU synchronization when creating observations of the environment. To avoid this I want to try using candle to reimplement the environment to keep everything on the GPU. Before I can do that I need to figure out whether I can send CUDA tensors from candle to a network implemented in tch-rs.

---

Related questions that you may also be able to answer: How would I send tensors between candle and pytorch+jax? What about tch-rs and pytorch+jax?

The issue names no files, tests, or entry points. Start by investigating the tensor interoperability APIs for candle, tch-rs, CUDA, PyTorch, and JAX; done would require a clearly documented, supported way to transfer tensors without GPU/CPU synchronization, if one is feasible.

Written by the indexing model from the issue text.