With the arrival of version 2.1 of Pytorch, Pytorch has defined new interfaces for extending custom hardware, which makes it more convenient to use custom hardware to leverage Pytorch,
GitHub - Ascend/pytorch: Ascend PyTorch adapter (torch_npu). Mirror of https://gitee.com/ascend/pytorch is a good example to use it, but it is too complicated,So does Pytorch have any plans to write a complete routine on how to adapt to custom hardware?
If you are still looking for something, this repo was helpful in explaining basic usage for registering a custom device (bdhirsh/pytorch_open_registration_example: Example of using pytorch’s open device registration API (github.com)).
My only issue with this example is that it doesn’t go into much detail about how to expand upon it. The code is well commented and explains what is being done, however I’m curious about how the developer (@bdhirsh) knows what to do (ie. it seems like there may be some documentation that I’m failing to find).