Public bug reported: ## FFe ##
Requesting a Feature Freeze exception for `migraphx`, in stonking, to land its first ever successfully-built version (10.0) in the archive. `migraphx` is now a member of the "rocm" packageset, but it was **not** yet a member at the time the coordinated stack-wide FFe for this transition (LP: #2164669, https://bugs.launchpad.net/ubuntu/+source/rocm/+bug/2164669) was filed and scoped to its 39 source packages, so it isn't covered by that bug. This is a follow-up/companion exception for the package that joined the set afterward, and should be tracked and landed alongside LP: #2164669. ### Scope: single source package (recently added to the rocm packageset) migraphx. Confirmed currently in the set via the live Launchpad packageset API: $ curl -s "https://api.launchpad.net/devel/package- sets/ubuntu/stonking/rocm?ws.op=getSourcesIncluded" returns 42 source packages, including `migraphx`, `hipfort`, and `rpp`, all three added to the set after #2164669 was filed against the original 39-package scope. This FFe covers `migraphx` only; `hipfort` and `rpp` are addressed separately if/as needed. ### Current state in the archive, every prior version failed to build `migraphx` has never had a successful build reach the release pocket: Date Status Target Pocket Component Section Version 2026-06-30 22:45:14 -03 Published Stonking proposed universe libdevel 7.1.1-7 2026-06-30 23:14:17 -03 Superseded Stonking proposed universe libdevel 7.1.1-6 2026-05-22 09:02:18 -03 Superseded Stonking proposed universe libdevel 7.1.1-5 2026-05-23 09:10:20 -03 Deleted Resolute proposed universe libdevel 7.1.1-5 2026-03-22 22:43:23 -03 Superseded Resolute proposed universe libdevel 7.0.2-1 All of these are Debian-derived 7.x uploads that failed to build and were superseded/deleted out of -proposed; none ever reached stonking's release pocket. In practice, no working `migraphx` has ever shipped to users. This request is to land the first one, directly at upstream 10.0. ### Description of the proposed changes New upstream version 10.0 (from 7.1.1), with the GPU backend actually enabled against the rocm 10.0 stack: - Compiler: switched from the distro `clang` to ROCm's own `clang-rocm` (`amdclang++`/`amdclang`), required for HIP device compilation - GPU backend on: MIOpen, rocBLAS, hipBLASLt, and Composable Kernel (CK), the last of which required and drove two upstream-facing fixes in `composable-kernel` itself (an `.install` file conflict between `libcomposable-kernel-dev`/`libcomposable-kernel-host-dev`, and a CMake exported-target bug where `hip::device` leaked HIP device-compile flags into any consumer); both fixed and verified end-to-end with a real migraphx build against the fixed composable-kernel packages - rocMLIR deliberately left OFF: it is not packaged for Debian/Ubuntu (upstream builds it from a pinned source commit at migraphx's own build time, vendoring a full copy of LLVM/MLIR's source, not a system package, and a DFSG concern if ever bundled). Documented as a known, intentional gap in `debian/rules`, not attempted here. - Debhelper compat bumped to 14 and associated packaging cleanups - Fixed several packaging gaps surfaced by the version bump: a missing `rocm_add_version_resource` CMake macro (not provided by Debian's rocm-cmake, guarded with a local patch), missing `.install` entries for upstream's new per-Python-version pybind11 module layout, a stale `libmigraphx-c3.symbols` file regenerated against the real 10.0 ABI, and two missing build-deps (`python3-sphinxcontrib.datatemplates`, `python3-dev`) that a clean chroot needs but a stale local build environment had masked ### Architecture impact, amd64 only (matches the rocm 10.0 stack) `migraphx` is restricted to `Architecture: amd64` (previous attempts also listed arm64, which never built anyway). This matches the rest of the rocm 10.0 stack, which is amd64-only on Ubuntu from this transition onward because it builds against AMD's LLVM fork (llvm-toolchain-rocm). Since no version of migraphx has ever successfully built or published on any architecture, this is not a drop relative to a working state, there is no working arm64 binary being removed. ### Reverse-dependency analysis, none $ reverse-depends src:migraphx Package not published in specified release $ reverse-depends src:migraphx -a source Package not published in specified release Nothing in the archive depends on migraphx, source or binary, expected, since no build of it has ever reached the release pocket. There is zero external regression risk from this package landing; the only question is whether the new package itself builds and works, which has been verified (see Testing performed). migraphx itself now build-depends on parts of the rocm 10.0 stack (`clang-rocm`, `hipcc-rocm`, `libamdhip64-dev`, `libhipblas-dev`, `libhipblaslt-dev`, `libhsa-runtime-dev`, `libmiopen-dev`, `librocblas-dev`, `libcomposable-kernel-host-dev`, `rocm-cmake`, `pkg-rocm-tools`, all `>= 10.0~`), so this upload is coupled to LP: #2164669 and should land alongside or after it, not before. ### Rationale for the exception Gets a previously perpetually-failing package into a working, publishable state for the first time, using the GPU backend it was actually designed to use (MIOpen/rocBLAS/hipBLASLt/Composable Kernel), rather than continuing to ship no working package at all. Deferring to a future cycle means another full release with no migraphx in the archive. ### Testing performed - Full builds succeeded, with the GPU backend enabled end-to-end, in `ppa:igorluppi/extra-10.0-exp`, including a build against the real, PPA-published (not hand-patched) fixed `composable-kernel` packages, confirmed via the Launchpad buildlog (`Status: successful`, amd64, ~78 minute build) - Verified no undefined symbols, no stray device-compilation leakage (migraphx has its own internal guard, `no_device.cpp`, that hard-errors if HIP device-compile flags leak into the wrong target, confirmed clean), and all 7 binary packages (`migraphx`, `migraphx-dev`, `migraphx-doc`, `migraphx-tools`, `migraphx-tests`, `libmigraphx-c3`, `libmigraphx-c-dev`) produced correctly - Also test-built and verified against a local container mirroring the target environment before each PPA upload ### Regression potential None to existing users, no working migraphx has ever shipped, so there is nothing currently in users' hands that this could regress. Risk is confined to whether the new package itself is sound, which has been verified through multiple successful builds as described above. ### Upstream changelog See https://github.com/ROCm/AMDMIGraphX/ (this package tracks AMDMIGraphX upstream releases directly). ** Affects: migraphx (Ubuntu) Importance: Undecided Status: New ** Summary changed: - FFe: migraphx — new package, first successful build at 10.0 (Ubuntu, amd64-only, joined the rocm packageset after LP: #2164669 was filed) + FFe: migraphx - new package, first successful build at 10.0 (Ubuntu, amd64-only, joined the rocm packageset after LP: #2164669 was filed) -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2168849 Title: FFe: migraphx - new package, first successful build at 10.0 (Ubuntu, amd64-only, joined the rocm packageset after LP: #2164669 was filed) To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/migraphx/+bug/2168849/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
