Bug#950306: nvidia-cuda-toolkit: Please package lecacy versions of cuda toolkit for unstable

[Andreas Beckmann]

Control: tag -1 wontfix

On 31/01/2020 09.14, Giacomo Mulas wrote:
> Would it be possible to introduce legacy versions of the cuda SDK in
> unstable, possibly using the alternatives system to make them coexist
> smoothly as done for the runtime?

The showstopper for providing older CUDA toolkit versions is their
dependency on ancient compilers no longer available in sid.
E.g. 9.2 and 10.0 require gcc-7 (or older), which is now gone from sid.

Andreas

Bug#950306: nvidia-cuda-toolkit: Please package lecacy versions of cuda toolkit for unstable

[Giacomo Mulas]

Package: nvidia-cuda-toolkit
Version: 10.1.168-6
Severity: wishlist

Dear Maintainer,

in recent times, nvidia has been dropping support faster and faster for
relatively old cards, that end up working only with legacy versions of the
drivers.  The Debian nvidia team has been doing an excellent work at making
most things work seamlessly via the selection system.  One notable exception
is the cuda toolkit: while one can run cuda executables on a machine with
the legacy drivers and cuda libraries (provided it does not require a cuda
level higher than the one supported by the legacy driver), the development
kit available in unstable is only version 10.x.  This means, for example
that I cannot anymore easily test/debug software on my laptop before running
on the production machine (in my case remote, hosted in an HPC centre).

Would it be possible to introduce legacy versions of the cuda SDK in
unstable, possibly using the alternatives system to make them coexist
smoothly as done for the runtime?

Thanks in advance, best regards
Giacomo Mulas