In the parent task we have installed manually rocm packages (http://repo.radeon.com/rocm/apt/debian/pool/main/) on stat1005, tested them with Tensorflow and the GPU that we bought. In order to do things properly, it is now the turn of importing those packages to wikimedia-buster.
There are some caveats:
- AMD provides packages only for Ubuntu 16.x and 18.x (https://rocm.github.io/ROCmInstall.html#supported-operating-systems---new-operating-systems-available). We tested them on Debian Buster and they seem to work fine.
- Most of the packages provided have also the source available, but we'd need to verify proper licensing of the software. Point of entry for all the repos is https://github.com/RadeonOpenCompute
- There is one package that contains binary libs that is not open source (yet). This package is in the dependencies of other ones that we need, and upstream seems not available to make it optional (https://github.com/RadeonOpenCompute/ROCm/issues/761). The package, IIUC, should be only used by image processing via OpenCL, something that is needed only in rare cases (non of the ones that we currently want to support like tensorflow). I manually removed the binary libs provided by the package and Tensorflow works as expected (as well as basic usage of OpenCL).
Some solutions that we could consider:
- import all the packages trying to follow-up/help upstream to finally remove the non open source dependency as soon as possible (so importing the non open source binary libs).
- import all the packages in boron, remove manually the non open source dependency from the related control files, rebuild all and import to wikimedia-buster.