nvidia-cuda-toolkit for buster

[Andreas Beckmann]

On 09/06/2023 16.27, Tobias Frost wrote:
> Hi,
> 
> I'm currently triaging nvidia-cuda-toolkit for buster:
> buster has 9.2.148-7+deb10u1, which is upstream version 9.2.148 with patch 1 [1]
> This seems to be the latest upstream version from the 9.2 series, and
> 9.2.x seem to be EOL, so there is no new upstream release expected to target any
> bugs or CVEs.

fixes are only available in the latest upstream release

> An option might be to update the cuda toolkit to a newer version. However:
> - Buster has nvidia drivers 418.xxx (EOL too). [2] tells us that this puts an upper
> cap on cuda-toolkit to version 10.1.

> - 10.1 are not binary-compatible with the 9.2 series, a SO-NAME bump is requred,
>    so cuda applications will not work without recompilation.

and probably require patching, too
(for the CUDA 11->12 transtion, 7 out of 9 packages requiring a rebuild 
do FTBFS with CUDA 12, I do not remember the numbers for 9.x/10.x/11, 
but at least since CUDA 11 the minor version transitions are gone)

> - 10.1 seems also to be EOL, but I couldn't find a site explictly stating so.
>    ([3] does not mention cuda versions <11 at all…)

There seems to be only one line of releases, no side branches at all.

> As said, higher version would require newer nvidia-drivers, and additionally,
> stop supporting several cards (e.g some Kepler cards, also on [3])

Despite of being documented, post-418 did not actually drop support for 
some card families. That finally happened again post-470.
But IIRC post-418 drops support for some non-glvnd gl/egl variants. 
(Probably not a big deal.)

So the 470 series might be an option, but I'm not sure if I'd really 
recommend that. This would limit us to CUDA 11.4.

But in case an update of the 418 driver is needed for supporting the lts 
kernel, the '418' branch is "ready" for that. Well, at least it will 
support build modules, but is otherwise untested.

> So my recommendation would be to put nvidia-cuda-toolkit on the "not-supported"
> list.

No objections.

Probably the same should be done for driver series that have reached EoL.
Currently there are only the 450 series (EoL 07/2023) and the 470 series 
(EoL 07/2024) (both LTSB branches), but there is not much point in using 
the 450 driver instead of the 470 driver as the latter supports a 
superset of the cards supported by the 450 series. (I'll turn the 450 
series into transitional packages depending on the 470 series after EoL; 
like I already did with the 460 series.)
So the "current" (470) driver series in bullseye will be EoL, too, 
around the time bullseye-lts support starts.

My plans for bookworm are to release with the 525 series (production 
branch, 1 year of support, EoL 12/2023) and switch to the next LTSB 
branch (3 years of support) once that comes (maybe switching to the next 
production branch inbetween). If another LTSB branch appears in time 
without dropping support for some cards, we might even switch again.
(530/535 seem to be new feature branches with no support plan (once 
superseded by a newer version).)

Andreas

PS: due to popular demand, I keep applying patches to the (long EoL) 340 
series in sid for supporting building against the latest kernels