Public bug reported:
[ Impact ]
* pytorch-rocm fails to build from source (FTBFS) in resolute once the ROCm
stack is updated to 7.2.4. Without this fix pytorch-rocm becomes
unbuildable in resolute after the ROCm 7.2.4 SRU lands, so it must be SRUed
together with the rest of the stack.
* Root cause: the ROCm 7.2.x stack builds against the LLVM 22 toolchain
(7.1.x used LLVM 21). pytorch-rocm's d/rules still pointed hipcc/clang at
the LLVM 21 device-library path, so the HIP device compile could not locate
the ROCm device bitcode:
clang++: error: cannot find ROCm device library; provide its path via
'--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build
without ROCm device library
* The fix is packaging-only. No change to pytorch itself — no source patch,
no ABI or API change on the pytorch side. It only realigns the build
toolchain (LLVM 21 -> 22) and the ROCm stack restriction (7.1 -> 7.2) so
the existing sources build against ROCm 7.2.4.
* Because there is no ABI/API change on the pytorch side, this update does not
interfere with future security fixes to pytorch-rocm — a later security
upload can be applied on top without conflicting with this packaging change.
And without this SRU pytorch-rocm being FTBFS would cause friction to apply
security fixes later.
[ Test Plan ]
* Reproduce: with the ROCm 7.2.4 stack available (from the staging PPA or
-proposed once the stack has landed), attempt to build the current resolute
pytorch-rocm (2.9.1+dfsg-1~exp2ubuntu3). The HIP device compile fails with
"cannot find ROCm device library" as above.
* Verify the fix: build the updated pytorch-rocm (llvm-22 /
HIP_DEVICE_LIB_PATH changes) against the ROCm 7.2.4 stack on amd64. The
package must build to completion and produce its binaries (done in the
staging PPA).
* Post-build sanity: install the resulting python3-torch (ROCm) on an AMD GPU
host and import torch:
python3 -c "import torch; print(torch.__version__);
print(torch.version.hip); print(torch.cuda.is_available())"
torch.version.hip should reflect the 7.2 build and a supported GPU detected.
[ Where problems could occur ]
* The change moves the whole pytorch-rocm build from LLVM 21 to LLVM 22. If
any part of pytorch's HIP codegen relied on LLVM-21-specific behaviour, it
could surface as a build failure or as miscompiled GPU kernels at runtime
rather than at build time. Regression testing should therefore include
actually running torch on GPU (a small tensor op / matmul), not just a
successful build.
* HIP_DEVICE_LIB_PATH is hard-coded to
/usr/lib/llvm-22/lib/clang/22/amdgcn/bitcode. If a future LLVM point update
changes that path layout, the device-lib lookup would break again the same
way — the same class of hard-coded path the bug fixes, just retargeted.
* This upload is inseparable from the ROCm 7.2.4 stack: building the new
pytorch-rocm against a 7.1.x/LLVM-21 stack would itself fail. It must
migrate together with the ROCm 7.2.4 SRU, not ahead of it.
* Risk is otherwise contained: no pytorch source code changed, so runtime
API behaviour of torch is unchanged relative to the current resolute
version apart from being built against the newer ROCm libraries.
[ Other Info ]
* Part of the ROCm 7.2.4 -> resolute stack SRU tracked in LP: #2150312
(src:rocm). pytorch-rocm is a reverse dependency of the ROCm stack that
FTBFS after the 7.2.4 bump; this bug is the packaging fix required to keep
it building.
** Affects: pytorch-rocm (Ubuntu)
Importance: Undecided
Status: Fix Committed
** Affects: pytorch-rocm (Ubuntu Resolute)
Importance: Undecided
Assignee: Talha Can Havadar (tchavadar)
Status: New
** Affects: pytorch-rocm (Ubuntu Stonking)
Importance: Undecided
Status: Fix Committed
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2161163
Title:
SRU: pytorch-rocm 2.9.1+dfsg-1~exp2ubuntu4 to resolute
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/pytorch-rocm/+bug/2161163/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
