rocm version via index url

[HPPinata]

--extra-index-url https://download.pytorch.org/whl/rocmX.Y is sufficient to install the right package.
This also makes A/B version testing easier.

[tazlin]

You did bring to my attention that the files did not match. However, you could practically get the same convenience that I think you're trying to bring to this file by using the main requirements.txt instead. It is supposed to be otherwise identical and I foresee that remaining to be the case.

The purpose of the rocm tag in the torch requirement is more to help avoid the pitfall of accidently installing the cpu or CUDA version of torch when using AMD. In fact, the existence of the requirements.rocm.txt file dates back to a time where you would have had to have had installed it manually via the command line.

There might be something I'm missing here. If you still think its worth adding, let me know why.

[HPPinata]

I just stumbled across this when trying to A/B test different ROCm versions.

With the explicit version tag removed this can be done exclusively in the Dockerfile building the environment.
If it's part of the requirements.txt I have to point the build to my fork.

I haven't looked at whether it "just works" with the standard requirements.txt but will test that shortly.

This isn't really too relevant (the fork exists now, might as well use it), just a bit annoying and a potentially unified requirements.txt might be a cleaner solution if it actually ends up working out.

[tazlin]

Yea, as it is, the rocm reqs file is more of a legacy artifact from when support was experimental. Maybe it should be unified sooner than later.