Hi Sebastian,
On 2025-12-01 14:25, Sebastian Ramacher wrote:
On 2025-11-28 15:02:10 -0700, Cordell Bloor wrote:
I assume this is the place where I should be requesting approval to upload
rocm-hipamd 5.7.1-7+deb14u1 to testing-proposed-updates. If I should be using
another format or sending it elsewhere, please let me know and I will make the
correction.
tpu is usually reserved issues that cannot or are hard to be solved via
unstable. This does not seem to be the case here. From the thread I
understand that the issue is fixed in rocm-hipamd in unstable -- but
please feel free to correct me if I understood that incorrectly. In that
case, we can schedule the autopkgtests with the right combination of
packages or we add the appropriate hints to ignore the autopkgtests. I
opted for the later.
Bug #1118618 exists in all versions of libamdhip64-5, but it is fixed in
libamdhip64-6 on unstable. It cannot be fixed in libamdhip64-5 via an
upload to unstable, but there will be no more libamdhip64-5 package on
testing once the transition is complete. The downside is that some
packages (such as librocsprase0, librocsprase0-tests, libhipsparse0,
and libhipsparse0-tests) are not going to exist with versions built
against libamdhip64-6, as they have been replaced
by librocsprase1, librocsprase1-tests, libhipsparse1,
and libhipsparse1-tests on unstable.
The src:rocm-hipamd on Trixie is practically identical to that on
Testing (aside from the changelog and version number) and it might not
be a bad idea to do the tpu as a dry run for an spu fixing the same bug
on Stable (to improve the Trixie -> Forky upgrade process and simplify
backports). With that said, it's certainly not required that we do the
tpu. We will have to complete the transition by removing some packages
regardless.
This leaves #1118939 that blocks migration to testing. What is the
status of this bug?
Bug #1118939 is really the same issue as Bug #1118618. The underlying
problem is that the libamdhip64-5 and newer libhsa-runtime-1 packages
are incompatible. One bug is filed against libamdhip64-5 and the other
against libhsa-runtime-1, but they're the same issue. That should
perhaps be reflected in the bug status... I'll update it.
The specific circumstance in the original report was fixed by the binNMU
for ucx that updated unstable to libucx0 (1.19.0+ds-1+b1). The libucx0
package is now built against libamdhip64-6 rather than libamdhip64-5.
However, I did not want to close the bug until we decided how we were
handling the incompatibility more generally.
Sincerely,
Cory Bloor
