10-26, 12:10–12:50 (Europe/Berlin), BoF Space 1 (meet under the @ sign in Cafeteria)
BoFs are sessions presented by community members as an opportunity to gather and discuss special topics of interest. BoFs can be anything from agenda-driven to an open-ended discussion.
Welcome, let's discuss CUDA-X full stack packaging for Debian, RPM, Nixpkgs, Conda, Spack, Julia, Conan, etc. ecosystems. Maintaining packages in the fast moving accelerated computing space can be a challenging labyrinth of incompatible dependencies and dlopen's.
For example: PyTorch with CUDA acceleration depends on TensorRT → cuDNN → cuBLAS & NVCC →CUDA runtime →NVIDIA driver
NVIDIA now provides "binary archives" tarballs and JSON manifests but what else can be improved in this space? How about a RSS feed (or similar) for new releases?
Another possible topic is that NVIDIA driver kernel modules are provided in two flavors: proprietary and GPL/MIT licensed, the latter allows redistribution of precompiled kmod packages. This has certain QoL improvements for end-users but it requires package maintainers to align many moving parts in sync - discussion is encouraged.
Guten Tag! I work @ NVIDIA on packaging for CUDA Toolkit, NVIDIA driver, cuDNN, cuQuantum, nvTIFF, etc. into various formats: Binary Archives (tarball), Conda, Debian, RPM, Runfile, Windows exe
At PackagingCon 2021, presented about Debian and RPM repository management.
In a past life, contributed Debian packages for the Maemo community and maintained an ArchLinux-based kiosk LiveUSB distro & AUR helper.