Hi Paul,
On 2026-10-15 10:39, Paul Gevers wrote:
On 2026-10-15 18:23, Cordell Bloor wrote:
So, this is not really a problem that is resolved by updating rocr-
runtime. The only thing we should do is remove that silly assertion
from src:rocm-hipamd on Trixie. We don't have to... but it would be
nice. I've been meaning to do that for ages. I'll make that my task
for today.
So are you saying that the issue doesn't affect unstable? Or are you
saying that without a fix in trixie this will be an issue during
upgrading to forky? This issue is blocking migration for months now.
It affects three rdepends on unstable (none of which are on testing).
The libamdhip64-6 transition was pretty messy (mea culpa). apt-rdepends
reports libamdhip64-5 has these reverse dependencies on unstable:
libamdhip64-5
Reverse Depends: librccl1 (>= 5.4.3-3+b1)
Reverse Depends: librccl1-tests (>= 5.4.3-3+b1)
Reverse Depends: libstdgpu-hip0d (>= 1.3.0+git20220507.32e0517-6+b2)
I need to address those. They're still depending on libamdhip64-5
because they FTBFS on unstable. The rccl problem would legitimately be
resolved by updating it to 7.2 from experimental (or any other version
that successfully builds). So, I suppose the ROCm 7.2 update would
resolve that part of the problem. I'm not sure about stdgpu, as it
seemed to have more problems than just ROCm.
With regards to the Trixie -> Forky upgrade, libhsa-runtime64-1 already
has a Breaks: libamdhip64-5 that will ensure the upgrade will succeed
(but may remove packages to do so). The only relevant rocr-runtime
change for this bug would be if libamdhip64-5 were fixed on Trixie and
finally removed entirely from Unstable. If that happens, we could put a
version limit on the Breaks: libamdhip64-5 in libhsa-runtime64-1,
thereby improving the experience for Trixie upgrades and backports.
Sincerely,
Cory Bloor
