** Description changed:

  [ Impact ]
  
   * This bug tracks the Stable Release Update of the whole ROCm 7.2.4 stack 
into
     Ubuntu 26.04 (resolute). resolute currently ships ROCm 7.1.0 (from the
     development cycle); 7.2.4 is the stable point release AMD recommends for
     this GPU-compute stack.
  
   * src:rocm is a metapackage only — it carries no code, it pulls in the ROCm
     component packages at a coherent version. SRUing src:rocm to 7.2.4 is the
     symbolic "top of the stack" marker that says "resolute now provides the
     ROCm 7.2.4 stack". The substantive work is in the individual component
     package SRUs listed under [ Other Info ]; each has its own Impact / Test
     Plan / regression analysis in its own bug.
  
   * User-facing effect: 7.2.4 brings the resolute ROCm stack up to the current
     stable upstream release (bugfixes, new-GPU enablement, updated math/comm
     libraries) instead of leaving 26.04 on the older 7.1.0 snapshot.
  
  [ Test Plan ]
  
   * Each component package SRU bug carries its own detailed test plan; this
     metapackage bug aggregates them. Do not re-run every component plan here —
     verify the stack installs and the per-package verifications pass. Only
     verification needed for this meta package is to run `apt install rocm` and
     see apt resolves 7.2.4 versions of the packages in rocm stack. This is also
     applicable for other binary packages published in this package.
  
   * The whole stack was rebuilt and co-installed in the staging PPA
     ppa:bullwinkle-team/rocm-7.2.4-resolute-sru before upload, to prove the
     component versions are mutually installable and build against each other.
  
   * Stack-level verification:
     - On a resolute amd64 host with -proposed enabled, install the metapackage
       (`apt install rocm`) and confirm it pulls the 7.2.4 component set with no
       unmet dependencies / no held-back packages.
     - `apt-cache policy rocm` shows the 7.2.4 metapackage from -proposed.
     - Smoke-test the runtime on AMD GPU hardware: `rocminfo` enumerates the
       agent(s), `clinfo` / a small HIP program runs.
  
   * Per-component functional verification (rocBLAS, rocFFT, MIOpen, RCCL, etc.)
     is tracked and signed off in the individual SRU bugs.
  
  [ Where problems could occur ]
  
   * This is a full-stack version bump, not a targeted cherry-pick, so the
     regression surface is the union of all component changes. The per-component
     bugs each carry their own "Where problems could occur" analysis; the
     stack-level risks are:
  
     - Version skew / co-installability: components are tightly inter-versioned.
       A component landing in -proposed out of order could leave the stack
       temporarily uninstallable. Mitigation: all components were co-built and
       co-installed in the staging PPA; they should migrate together.
  
     - Toolchain: 7.2.x builds against the ROCm LLVM toolchain (22); a mismatch
       between the LLVM/clang version a component expects and what is available
       in resolute would surface as build or HIPRTC-at-runtime failures.
  
   * Because this is a metapackage, a regression will not manifest in src:rocm
     itself — it will surface in one of the component packages or in a reverse
     dependency at install or runtime.
  
  [ Other Info ]
  
   == Component SRU bugs (ROCm 7.2.4 → resolute) ==
  
   Core / runtime / toolchain:
     - rocm-core           LP: #2153989
     - rocm-cmake          LP: #2153804
     - rocm-llvm           LP: #2153424
     - rocm-hipamd         LP: #2156375
     - rocr-runtime        LP: #2153419
     - rocdbgapi           LP: #2153818
     - rocminfo            LP: #2153792
     - rocm-smi-lib        LP: #2154039
     - amdsmi              LP: #2153816
     - roctracer           LP: #2153991
     - hipify              LP: #2153789
     - pkg-rocm-tools      LP: #2159601   (gfx950 / MI350 CDNA4 enablement)
  
   Math / comm libraries:
     - rocblas             LP: #2155118
     - hipblas             LP: #2156532
     - hipblas-common      LP: #2154225
     - hipblaslt           LP: #2155653
     - rocsparse           LP: #2154553
     - hipsparse           LP: #2156252
     - rocsolver           LP: #2155138
     - hipsolver           LP: #2156046
     - rocfft              LP: #2155392
     - hipfft              LP: #2155398
     - rocrand             LP: #2153990
     - hiprand             LP: #2154493
     - rocprim             LP: #2154360
     - rocthrust           LP: #2154711
     - hipcub              LP: #2155648
     - rocalution          LP: #2154350
     - rccl                LP: #2154290
     - rocwmma             LP: #2156085
     - miopen              LP: #2157988
+    - pytorch-rocm        LP: #2161163
  
   Notes:
     - llvm-toolchain-rocm is superseded by rocm-llvm for the 7.2.4 stack; the
       old 7.1.0 source is not being SRUed.
     - magma-rocm is a reverse dependency, not a ROCm stack component; tracked
       separately (see rebuild status below).
  
   == Staging PPA build status ==
  
   Built and tested in ppa:bullwinkle-team/rocm-7.2.4-resolute-sru. All 7.2.4
   stack components are Published for amd64/amd64v3/arm64/ppc64el(for some).
   The build failures /  dependency-waits observed in the PPA fall into two
   classes, none of which block the SRU:
  
   (1) Pre-existing / not-our-regression reverse-dependency failures:
       Reverse dependencies that FTBFS for reasons unrelated to the ROCm update:
         - ucx (amd64): FTBFS at config/parser.c:649:11
           "error: assignment discards 'const' qualifier ...
           [-Werror=discarded-qualifiers]" — a compiler-strictness failure in
           ucx's own source, independent of the ROCm bump.
         - haskell-futhark (ppc64el), slurm-wlm-contrib (armhf/ppc64el),
           papi (s390x), spfft (s390x): failures on non-primary architectures /
           unrelated toolchain issues.
  
   (2) Dependency-waits noise:
       Dependency-waits and failures on armhf / ppc64el / riscv64 / s390x for
       atlas-ecmwf, papi, starpu-contrib, ucc, slurm-wlm-contrib, rocalution
       (armhf). ROCm targets amd64 (and arm64 where supported); these arches are
       not release-relevant for the ROCm stack.
  
  Please also see https://canonical.github.io/rocm-qa/dependency-graph for
  the dependency relationship between the packages in ROCm stack.
  
  =========== Initial Request by Pander ================
  Please upgrade rocm to at least version 7.2.2
  
  See also https://rocm.docs.amd.com/projects/install-on-
  linux/en/latest/install/quick-start.html#
  
  The currently packaged version for resolute is 7.1.0 is from Oct 30,
  2025.

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2150312

Title:
  SRU: ROCm Stack 7.2.4 to Ubuntu Resolute

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/rocm/+bug/2150312/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to