This bug was fixed in the package ggml - 0.9.7-2ubuntu1
---------------
ggml (0.9.7-2ubuntu1) resolute; urgency=medium
* Merge with Debian unstable. Remaining changes:
- d/control: make sure to use hipblas compatible with ggml 0.9.6
* Dropped:
- d/p/force-no-lto-for-cpu-feats.patch: fix ppc64el fail with lto
[In 0.9.7-2]
ggml (0.9.7-2) unstable; urgency=medium
[ Christian Kastner ]
* Bump copyrights
* Update patch metadata
[ Talha Can Havadar ]
* d/p/force-no-lto-for-cpu-feats.patch: fix ppc64el fail with lto
ggml (0.9.7-1ubuntu1) resolute; urgency=medium
* Merge with Debian unstable. Remaining changes:
- d/control: make sure to use hipblas compatible with ggml 0.9.6
- d/p/force-no-lto-for-cpu-feats.patch: fix ppc64el fail with lto
When LTO enabled in build environments it forces all builds to have LTO
in place. But feature detection logic is fragile, and causing Illegal
instruction errors with lto. This disables LTO for the feature
detection code to prevent cross-module optimization from inlining
architecture-specific instructions into the score function. Without this,
LTO can cause SIGILL when loading backends on older CPUs (e.g., loading
power10 backend on power9 crashes before feature check runs).
ggml (0.9.7-1) unstable; urgency=medium
* New upstream version 0.9.7
ggml (0.9.6-4ubuntu1) resolute; urgency=medium
* d/p/force-no-lto-for-cpu-feats.patch: fix ppc64el fail with lto.
When LTO enabled in build environments it forces all builds to have LTO
in place. But feature detection logic is fragile, and causing Illegal
instruction errors with lto. This disables LTO for the feature
detection code to prevent cross-module optimization from inlining
architecture-specific instructions into the score function. Without this,
LTO can cause SIGILL when loading backends on older CPUs (e.g., loading
power10 backend on power9 crashes before feature check runs). (LP: #2138963)
* d/control: make sure to use hipblas compatible with ggml 0.9.6
ggml (0.9.6-4) unstable; urgency=medium
* CUDA backend: drop ppc64el as a supported architecture
ppc64el is no longer a CUDA architecture.
* CUDA backend: Fix build number
* Let libggml0-backend-vulkan suggest drivers
* Add helper for running tests on GPUs
* autopkgtest:
- Remove unnecessary indirections
- Re-enable HIP tests
- Add CUDA tests
- Add Vulkan tests
- Simplify test names
ggml (0.9.6-3) unstable; urgency=medium
* Add another workaround for ROCm on trixie
ggml (0.9.6-2) unstable; urgency=medium
* libggml0.symbols: Add new symbols from 0.9.6
ggml (0.9.6-1) unstable; urgency=medium
* New upstream version 0.9.6
* Refresh patches
ggml (0.9.5-6) unstable; urgency=medium
* Update ROCm downgrade patch to fix init in trixie
* Add patch debian/patches/Use-only-CPU-backend-in-build-time-tests.patch
ggml (0.9.5-5) unstable; urgency=medium
* More updates to the ROCm 5.5 downgrade patch
Fix a FTBFS with gfx110n as a target
ggml (0.9.5-4) unstable; urgency=medium
* Drop Priority field overlooked in previous release
* More updates to the ROCm 5.5 downgrade patch
ggml (0.9.5-3) unstable; urgency=medium
* CUDA: Adapt not-installed list to new library locations
* Bump Standards-Version to 4.7.3
- Drop Priority: optional, it's the dpkg default since trixie
* Bump copyrights
ggml (0.9.5-2) unstable; urgency=medium
* Temporarily revert riscv64 to minimal feature flags (Closes:
#1124644)
ggml (0.9.5-1) unstable; urgency=medium
[ Christian Kastner ]
* Create d/NEWS, announcing the move to a public library
* Drop RPATH from now public libraries
* Drop patches obsoleted by upstream introduction of SOVERSION
* Update package descriptions, dropping the hint to private
* Lower dependency on libamdhip64-dev to the fixed deb13u1 version
* Re-enable dh_makeslibs
* Start tracking symbols in d/libggml0.symbols
* Switch riscv64 to runtime feature detection
* d/copyright: Drop the hint to private
* Set GGML_BUILD_VERSION to Debian revision
* Refresh d/copyright
[ Mathieu Baudier ]
* New upstream version 0.9.5.
* Install libraries and headers to public paths. (Closes: #1117476)
ggml (0.9.4-7) unstable; urgency=medium
* Re-enable the HIP backend
* Merge libggml0-backend-cpu into libggml0
* Drop CUDA backend, now built by src:ggml-cuda
* Re-incorporate src:ggml -> src:ggml-cuda transformation
* Fix backend version enforcment
The version on the Suggests is ineffective. We need to enforce version
equality through a Depends on the backends, which overrides the
dependency that ${shlibs:Depends} will generate.
* Suggest libggml0-backend-cuda
* Enable build-time tests
* Salsa CI: Use pkg.ggml.minhip build profile
ggml (0.9.4-6) unstable; urgency=medium
* Move pkgconfig to multi-arch aware location.
This was an oversight when in the Great Reorganization of the llama.cpp
and ggml packages, leading to a multi-arch file conflict.
ggml (0.9.4-5) unstable; urgency=medium
* Also disable autopkgtests for temporarily disabled HIP backend
ggml (0.9.4-4) unstable; urgency=medium
[ Christian Kastner ]
* Temporarily disable the HIP backend in unstable, until the ROCm stack has
stabilized again
* Introduce a build profile for single AMD GPU target, to enable reducing
Salsa load
[ Cordell Bloor ]
* Drop HIP backend for ppc64el
* Use default CMAKE_HIP_COMPILER as detected by CMake
ggml (0.9.4-3) unstable; urgency=medium
* d/rules: Don't hardcode clang version. Resolves a FTBFS
ggml (0.9.4-2) unstable; urgency=medium
[ Aurelien Jarno ]
* Disable zicbop and zfh on riscv64.
They are not part of our baseline. (Closes: #1117633)
ggml (0.9.4-1) unstable; urgency=medium
* New upstream version 0.9.4. (Closes: #1116954)
- Drop patches included upstream
- Refresh patches
* Convert d/watch to version 5
* Drop hwcaps from private directories patch
ggml (0.0~git20250712.d62df60-5) unstable; urgency=medium
* For architectures that don't yet support GGML_CPU_ALL_VARIANTS, disable
all features not supported by our baseline spec.
Thanks to Aurelien for the pointer. (Closes: #1114747)
ggml (0.0~git20250712.d62df60-4) unstable; urgency=medium
* Re-enable loong64, ppc64, riscv64, s390x builds.
These don't support GGML_CPU_ALL_VARIANTS yet, so we build them without
any flags enabled. Performance will suffer. I intend to submit fixes for
this upstream.
* Bump pkg-rocm-tools dependency
ggml (0.0~git20250712.d62df60-3) unstable; urgency=medium
[ Otto Kekäläinen ]
Disable the two permanently failing jobs on Salsa CI
[ Christian Kastner ]
* Limit architectures to those with ALL_VARIANTS support.
Support for loong64, riscv64, and s390x will be added soon
* Salsa: Correct typo in DEB_BUILD_PROFILES
* Salsa: Build the HIP package only for one GPU, to reduce build time
* autopkgtest: Correct test dependency names
ggml (0.0~git20250712.d62df60-2) unstable; urgency=medium
* Upload to unstable
* Fix/improve package descriptions
* Make GPU-specific backends conflict with Vulkan backend
* Use rocm-target-arch to select and document build GPU arch
ggml (0.0~git20250712.d62df60-2~exp2) experimental; urgency=medium
* CUDA backend is in section contrib/libs
* Extend Breaks/Replaces to libggml-dev
ggml (0.0~git20250712.d62df60-2~exp1) experimental; urgency=medium
* Switch build to GGML_BACKEND_DL=ON (backends as plugins)
* Add SOVER, and rename packages to match SONAME and plugin status
* Build the CUDA backend via Build-Profiles
* Add Add-soversion-to-libraries.patch
* Generate GGML_BUILD_NUMBER based on upstream date
* Change Section field to 'libs'
* Factor out a -dev package
* New layout Breaks/Replaces old layout
* d/rules: Don't create ldconfig triggers for private libraries
* Salsa-CI: disable HIP build
-- Talha Can Havadar <[email protected]> Wed, 18 Feb
2026 22:12:55 +0100
** Changed in: ggml (Ubuntu)
Status: In Progress => Fix Released
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2138963
Title:
ggml 0.9.5-3 ppc64el build failure during testing phase
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/ggml/+bug/2138963/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
