It is said that, the newest ROCm version, 5.6.0 Alpha, supports some AMD consumer GPUs on Windows now.
Would Pytorch be supporting AMD GPUs on Windows soon?
I am also interested. Here is the article if anyone is interested in reading: AMD ROCm Comes To Windows On Consumer GPUs | Tom's Hardware
Part of the AMD ROCm Platform, the HIP SDK is now available on Windows. 5.5 was released on July 27th, 2023.
now it’s online: https://www.amd.com/en/developer/rocm-hub/hip-sdk.html
GPU and OS Support (Windows) — ROCm Documentation Home
Yes, I am very interested too about this, and hope it is supported soon with pytorch, so we can get the most out of the newer AMD cards.
Being curious and impatient, I installed the latest ROCm windows SDK and grabbed the source to build pytorch for Windows and ROCm, this is what I found.
There are a lot of assumptions in the build files and in the code made around which libraries are available on which platforms, eg: CUDA, only if NOT_MAC, ROCm must be on linux. that is one thing that will require some maintenance but that’s not so bad.
There was one place in code that would be used on windows that wasn’t where a thread exposes a bool that needs changed to get it to build.
That’s about it, although I got it to build, there was a lot of those assumptions with definitions and the package I produced resulted in a “working” pytorch that spewed errors about CUDA not being available if you tried to do any heavy lifting with it. This satisfied my curiosity about if the ROCm windows SDK even had all the math libraries one would need to build pytorch and it appears that it does. It shouldn’t be far off that someone familiar with this project could pick off this low hanging fruit.