ci.rocm.debian.net: a Debian CI with AMD GPUs, for AI/ML packages

Speaker: Christian Kastner

Track: Artificial intelligence & Debian

Type: Long talk (45 minutes)

Room: Grand amphi

Time: Jul 17 (Thu): 10:30

Duration: 0:45

Much of number-crunching today is performed on GPUs. This especially applies to AI/ML-driven workloads. ROCm is AMD’s open source software stack for general purpose GPU computing, challenging Nvidia’s CUDA.

The Debian ROCm team’s mission is to integrate and support ROCm on Debian to the point that using an AMD GPU (enterprise and consumer) for computation tasks is only an apt-get away. And with “support”, we don’t just mean packaging the libraries: we want to actually demonstrate just how well ROCm works.

To this end, we forked Debian’s CI tools and bootstrapped our own environment with GPU support, currently featuring more than a dozen GPU architectures. Our CI works exactly like Debian’s official CI, except that we only look at the subset of packages linked to ROCm libraries.

In this talk, I’ll present a brief overview of our CI, and the policy and technical challenges we faced bootstrapping it. I’ll then show how contributors can adapt their packages to get coverage in our CI.

URLs