Hello,The packaging of rocRAND is a nice milestone, as it is the first library packaged for Debian that was written in the HIP language. Étienne has been doing some great work! Its submission into new marks the start of a new stage of ROCm packaging. I thought this might therefore be a good time to look at what still lies ahead as we work towards some long-term goals. Given that this is the Debian AI mailing list, I imagine there is probably significant interest in enabling AMD GPU support in PyTorch and Tensorflow packages. I've therefore put together a couple diagrams that illustrate the remaining packages required to reach that objective.
In the attached diagrams, each box represents a tool or library to be packaged. The arrows point towards the package dependencies. I put a suggested order to approach packaging in the upper-right corner and an estimated difficulty on the bottom (with four stars being the maximum difficulty). PyTorch and Tensorflow depend on nearly all of the ROCm math libraries, so there's roughly twenty libraries to package.
The combined diagram for both PyTorch and Tensorflow is rather busy, so I've also attached separate diagrams for each. They were generated from a dot file, which is also attached.
Sincerely, Cory Bloor
Attachment:
pytorch-tensorflow-order.png
Description: PNG image
Attachment:
tensorflow-order.png
Description: PNG image
Attachment:
pytorch-order.png
Description: PNG image
Attachment:
rocm-deptree.dot
Description: application/msword-template
