Source: autopkgtest
Version: 5.43
Severity: wishlist
Dear autopkgtest maintainers,
(re-submitting as a bug, as my original mail to autopkgtest@package.d.o
probably didn't surface where I expected it to)
pkg-rocm-tools [1] is (resp: was, accepted today) sitting in NEW and
contains a feature that I only now realize I probably should have
cleared beforehand with you.
Intro: this package contains a collection of scripts that facilitate
using AMD GPUs in rootless containers and VMs.
To this end, it also ships:
/usr/bin/autopkgtest-virt-podman+rocm
/usr/bin/autopkgtest-virt-qemu+rocm
These are simple wrappers around the respective autopkgtest backends.
The wrappers add a few options, take care of setup, etc. For example,
the following will attach the GPU on PCI slot 03:00.0 to the VM, and
perform various checks (is it assigned to VFIO for pass-through, does
the user have the right permissions, etc.):
$ autopkgtest -B rocrand -- qemu+rocm --gpu 03:00.0
unstable-autopkgtest-amd64.img
I uploaded pkg-rocm-tools to main as we consider it stable, it is
generally useful, and most importantly has a helper script for any other
Debian contributor to easily run autopkgtests against AMD GPUs in our
infra. This has been requested a few times now.
But having used this naturally for so long, it didn't occur to me in
time that I should have checked whether shipping this in effectively the
autopkgtest namespace would be OK with you.
I would appreciate your feedback on the matter. Is this approach
acceptable to you? If not, can you suggest another approach to which
you would be more amenable?
I'm leaving the package in experimental, pending resolution.
Best,
Christian
PS: We've been using this for more than a year on our own CI [2] which
runs a number of forked and new components [3].
[1]: https://salsa.debian.org/rocm-team/pkg-rocm-tools
[2]: https://ci.rocm.debian.net/
[3]:
https://salsa.debian.org/rocm-team/rocm-team-infra/-/blob/master/doc/Software-in-our-infra.md
