Your message dated Sat, 6 Dec 2025 17:20:26 +0100
with message-id <[email protected]>
and subject line Re: Bug#1118212: transition: rocm-hipamd
has caused the Debian Bug report #1118212,
regarding transition: rocm-hipamd
to be marked as done.
This means that you claim that the problem has been dealt with.
If this is not the case it is now your responsibility to reopen the
Bug report if necessary, and/or fix the problem forthwith.
(NB: If you are a system administrator and have no idea what this
message is talking about, this may indicate a serious mail system
misconfiguration somewhere. Please contact [email protected]
immediately.)
--
1118212: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1118212
Debian Bug Tracking System
Contact [email protected] with problems
--- Begin Message ---
Package: release.debian.org
Severity: normal
X-Debbugs-Cc: [email protected], [email protected]
Control: affects -1 + src:rocm-hipamd
User: [email protected]
Usertags: transition
I am requesting a transition slot for the ROCm GPU compute stack. The
key package of this transition is libamdhip64-5. The transition is both a
change ABI change and a switch from the LLVM 17 to LLVM 21 compiler for
the HIP language. The update to the runtime will likely force the update
to a number of other components in the ROCm stack. I would suggest that
we transition the whole ROCm stack from 5.7 to 6.4. This would mean transitions
for libcomgr2, librocblas0, libhipblas0, librocsparse1, and libhipsparse1
as well.
The Debian ROCm Team does not have particularly good systems to handle
this transition at the moment (especially with regards to the compiler).
That is something that I hope to remedy in the future. At the moment, there
are a number of reverse dependencies that explicitly list LLVM 17 components
in their d/control or d/rules files. I expect that most or all of these
components will FTBFS following the transition.
Outside of packages maintained by the Debian ROCm Team, there are only a
few affected reverse dependencies. Those would be: magma-rocm, ggml,
haskell-futhark, mpich, slurm-wlm, spfft, starpu, stdgpu, and ucx. I
will be working through these packages one-by-one and providing NMUs.
The version of ROCm on Debian Sid is ancient at this point and although
the transition might cause temporary runtime breakages to AMD GPU compute
packages on unstable, I believe it would be best to begin rolling forward.
We're already seeing key parts of the ROCm stack (such as rocblas) getting
removed from testing due to the dependency on LLVM 17.
Sincerely,
Cory Bloor
Ben file:
title = "rocm-hipamd";
is_affected = .depends ~ "libamdhip64-5" | .depends ~ "libhiprtc-builtins5" |
.depends ~ "libamdhip64-6" | .depends ~ "libhiprtc-builtins6";
is_good = .depends ~ "libamdhip64-6" | .depends ~ "libhiprtc-builtins6";
is_bad = .depends ~ "libamdhip64-5" | .depends ~ "libhiprtc-builtins5";
--- End Message ---
--- Begin Message ---
On 28/11/2025 23:02, 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.
[ Reason ]
This update fixes #1118618, which is an incompatibility between newer
versions of libhsa-runtime-1 and libamdhip64-5, as well as between
libhsa-runtime-1 and rocm-opencl-icd. Newer versions of libhsa-runtime-1
reflect this by including a Breaks declaration for libamdhip64-5 and
older versions of rocm-opencl-icd, but this is making the transition to
libamdhip64-6 difficult.
This fix to libamdhip64-5 would allow the upload of a new version of
libhsa-runtime-1
with a maximum version on the breaks declaration, thereby allowing
libhsa-runtime-1
to migrate to testing and unblocking src:rocm-hipamd.
This update also fixes missing dependencies in the rocm-hipamd autopkgtests.
The autopkgtests invoke clang++-17 but the clang-17 package was only installed
because hipcc depended on it. The rocm-hipamd autopkgtests thus begun to fail
when hipcc updated its dependency on clang-17 to clang-21. The clang-17
functionality being tested by the rocm-hipamd autopkgtests works even after
hipcc is updated to clang-21, as long as clang-17 is actually installed.
[ Impact ]
The libhsa-runtime-1, libamdhip64-5 and rocm-opencl-icd packages are
dependencies
of nearly all packages on testing maintained by the Debian ROCm Team. The
incompatibility makes it difficult to compare old and new versions of packages,
and has made the transition very disruptive.
[ Tests ]
I've used librocsparse1-tests as the test for libamdhip64-5 and clpeak as the
test
for rocm-opencl-icd. I've attached a test log (edited for length) that shows
that
the updated src:rocm-hipamd package works on testing with existing package
versions,
will continue to work on testing once libhsa-runtime-1 updates its breaks
annotations
in a new upload (planned for src:rocr-runtime 6.4.3+dfsg-5).
I then do a quick sanity check showing that these tests crashed when using the
existing
version of libamdhip64-5/rocm-opencl-icd on testing together with
libhsa-runtime-1 6.4.3+dfsg
from before the breaks annotations were added. Then, I update only
libamdhip64-5 and
rocm-opencl-icd with this proposed change to show that this change fixes the
original
justification for the breakage.
[ Risks ]
This change is small and the package version will quickly be replaced when
the src:rocm-hipamd transition is completed.
As shown by the attached debdiff, this change is the removal of an unnecessary
check
(using a patch backported from upstream) and the explicit declaration a
dependency
on clang-17 packages added to the d/t/control file.
[ Checklist ]
[x]*all* changes are documented in the d/changelog
[x] I reviewed all changes and I approve them
[x] attach debdiff against the package in testing
[x] the issue is verified as fixed in unstable
[ Changes ]
* Add clang-17 and rocm-device-libs-17 to Depends in d/t/control
* Add d/p/0042-support-generic-isas.patch to remove a check that the current
device only supports one ISA
I was looking at this today, but the rocm stack has finally migrated (thanks to
Sebastian's hints). So maybe that fix will need to go to trixie-pu, but for the
transition purposes there's nothing further to do.
I have also been able to remove rocm-device-libs and rocm-compilersupport from
testing (#1084072) now that rocm-llvm has migrated.
Cheers,
Emilio
--- End Message ---
