Hi Paul,
On 2026-08-09 08:10, Paul Gevers wrote:
> While I haven't come back to you on this, I challenged you last year (at
> DebConf) to work towards integration of your infrastructure to tests
> GPU's with the official ci.d.n infrastructure such that we don't need to
> skip tests like these. Did you make any progress on that? There's quite
> some work ongoing this year to change our setup, so maybe now would be a
> good moment to discuss this again. (It dropped a bit off my radar, but
> your comment reminded me).
there's been quite a bit of progress actually, I just haven't found the
time to finalize it yet.
The two most important changes since our discussion:
- the CI I am working on has been made vendor-neutral
The ROCm CI presented at DebConf25 was tied to AMD. I've abstracted
the vendor away, and the new CI (and tools) all support AMD and
NVIDIA, and new vendors can be added quickly.
- the CI also supports non-free data
For example, src:llama.cpp's tests requires non-free LLM models
to test, which are passed to the test bed at test time
Some of the features needed to support this have already been upstreamed
into autopkgtest and debci, some I still need to file. Some extra
tooling, I still need to upload to the official Archive.
This was all actually driven by the wish to make src:llama.cpp testable,
of which I am one of the maintainers. llama.cpp is a very rapidly
evolving software, and reflects typical AI/ML use cases well. With
upstream's development pace, having a CI to test new releases on was vital.
I think this prototype CI is pretty much done, and discussion should
happen soon. I'll be doing a write-up on that in ~September. I'd really
like to see this be live by DeConf27.
Best,
Christian
