I've been working on a project called Wooly for a while now and have a solid beta for it. It lets you run your Pytorch projects from any CPU machine and get GPU speeds. Instead of hard-partitioning GPUs or overpaying for unused GPU reserved/hold time, Wooly manages execution of multiple user workloads over the network on our remote GPUs.

We've got a free beta available for you to try. Check it out at https://docs.woolyai.com