Bug#1118212: transition: rocm-hipamd
Hi Emilio,
I'm sorry this was such a mess. The ROCm stack is complex and I've
certainly made a number of mistakes in this transition. Thank you for
your help in moving this forward.
On Tue, 25 Nov 2025 08:53:38 +0100 Emilio Pozuelo Monfort
<pochu@debian.org> wrote:
> On 25/11/2025 00:53, Cordell Bloor wrote:
> > I verified that the rocm-hipamd autopkgtests could be fixed by adding
> > clang-17 to the test dependencies [1], but I wasn't able to upload that
> > fix because rocm-hipamd 5.7.1-7 would FTBFS on unstable. My guess is
> > that we should outright remove rocm-hipamd from testing and file a bug
> ^^^^^^^^^^^
> Do you mean removing rocm-llvm from testing?
>
> In any case, if this is a test-only issue, we could add a hint to
ignore it. Or
> if it is in rocm-hipamd but you can fix it in sid, rocm-llvm could add
> appropriate versioned depends or breaks in order to be tested against
> rocm-hipamd/sid.
No. The src:rocm-llvm package doesn't exist in testing. I did mean
src:rocm-hipamd. It is a test-only problem in src:rocm-hipamd.
I'd normally prefer not to change a package relationship to work around
a test-only problem, but I don't think it will matter in practice so
I've added a libhipamd64-dev (>= 6.4.3~) version requirement to hipcc in
src:rocm-llvm 7.0.2+dfsg-2.
> > Good point. There's a deprecation warning causing the autopkgtest to
> > fail. I didn't mention it, but I have a fix pending upload [2].
That's a
> > blocker for migration, so I will prioritize that upload ASAP.
>
> Cool, thanks.
I've uploaded src:rocm-hipamd 6.4.3-4 with the fix for the deprecation
warnings in the autopkgtests, but I'm afraid that there are a number of
new failures. The problem is that libamdhip64-6 from src:rocm-hipamd
6.4.3-4 requires libhsa-runtime64-1 from src:rocr-runtime 6.4.3+dfsg-4.
However, libhsa-runtime64-1 6.4.3+dfsg-4 breaks libamdhip64-5 [1], and
libamdhip64-6 requires libhsa-runtime64-1 6.4.3+dfsg-4. This passed the
first time because the breaks relationship hadn't been added yet.
The breaks relationship between libhsa-runtime64-1 and libamdhip64-5 is
(IMO) due to a bug in the latter. The libhsa-runtime64-1 library returns
a list of compatible GPU ISAs for the current device and libamdhip64-5
asserts that there is only one item in the list. Newer versions of
libhsa-runtime64-1 return a list with two items, which results in
libamdhip64-5 returning an error. This could /probably/ be fixed by
patching libamdhip64-5 to ignore the second item in the list, but such a
fix would have to be uploaded directly to testing. My feeling is that
this is not worth doing, but I'm willing to do so if needed.
The next question is why is the debci running autopkgtests for libraries
that depend on libamdhip64-5? I suspect that this is because
src:rocsparse, src:hipsparse, and src:rocrand have all had package removals:
src:rocrand has dropped libhiprand-dev, libhiprand-doc, libhiprand1, and
libhiprand1-tests. These are to be provided by src:hiprand, which is
currently in NEW [2].
src:rocsparse is transitioning from librocsparse0 to librocsparse1
src:hipsparse is transitioning from libhipsparse0 to libhipsparse1
In retrospect, perhaps I have done separate formal transitions for
rocsparse and hipsparse. I guess that is how they ultimately get removed
from testing. In any case, it looks to me like those transitions are
complete on unstable [3][4].
I suppose I need to file RM requests for testing to remove binaries for
rocm-hipamd (due to libhipamd-doc), rocrand (due to libhiprand*),
rocsparse (in lieu of its formal transition), and hipsparse (ditto)?
All I can say is thank you for your patience. This has been quite a
learning experience, and I promise I will apply the lessons learned from
this transition to do a better job on the ROCm 6 -> ROCm 7 transition
next year.
Sincerely,
Cory Bloor
[1]: https://bugs.debian.org/1118618
[2]: https://ftp-master.debian.org/new/hiprand_6.4.3-1~exp1.html
[3]: https://release.debian.org/transitions/html/auto-rocsparse.html
[4]: https://release.debian.org/transitions/html/auto-hipsparse.html
Reply to: