Public bug reported:

[ Impact ]

 * This update bumps hipBLASLt from 7.1.1 to 7.2.4 as part of the
   coordinated ROCm 7.2 stack release. The new upstream version delivers
   expanded hardware and data-type support, a correctness fix, and an
   enumeration update:

   - Support for the BF16 data type (and BF16 input with FP32 output)
     on gfx90a (MI200/MI210/MI250). Users running mixed-precision workloads
     on Aldebaran hardware could not use BF16 with hipBLASLt in ROCm 7.1.
   - hipBLASLtExt operation APIs are now supported on gfx11XX and gfx12XX
     (RDNA3 and RDNA4 GPUs), extending the reach of the extension API to
     consumer and workstation hardware.
   - Sigmoid Activation Function is added as a new epilogue option, with a
     follow-up enumeration value fix (ROCm 7.2.1) ensuring the activation
     variant is correctly identified by downstream code.
   - Added HIPBLASLT_OVERRIDE_COMPUTE_TYPE_XF32 environment variable,
     allowing users to override the XF32 compute type when precision
     trade-offs are desired.
   - Fixed: users could encounter HIPBLAS_STATUS_INTERNAL_ERROR with various
     matrix sizes in CPX (complex) mode. This caused silent failures in
     complex-valued GEMM operations.

   hipBLASLt is a dependency of PyTorch-ROCm and other ML frameworks for
   GEMM acceleration on AMD GPUs.

[ Test Plan ]

1. Build:
   - dpkg-buildpackage or sbuild succeeds.
   - dpkg --compare-versions 7.2.4+dfsg-0ubuntu1 gt 7.1.1+dfsg-3ubuntu2
     returns true.
   - No debian/*.symbols file exists for this package; 
DPKG_GENSYMBOLS_CHECK_LEVEL=4
     is set in debian/rules and will flag any ABI changes at build time.

2. Installability:
   - apt install libhipblaslt1 libhipblaslt-dev succeeds.
   - Confirm reverse dependencies (e.g. PyTorch-ROCm) remain installable
     without rebuild.

3. Autopkgtest (hipblaslt1-tests) on a GPU-equipped testbed:

autopkgtest [22:22:40]: starting date and time: 2026-06-04 22:22:40+0000
autopkgtest [22:22:40]: version 5.55
autopkgtest [22:22:40]: host rocmtest; command line: /usr/bin/autopkgtest 
hipblaslt -U --timeout-test=28800 '--setup-commands=export NEEDRESTART_MODE=a; 
export NEEDRESTART_SUSPEND=1' '--setup-commands=add-apt-repository -ss -y' 
'--setup-commands=add-apt-repository ppa:igorluppi/hipblaslt-7.2.4 
--enable-source -y' -- lxd ubuntu-daily:resolute --profile=rocm-gpu
autopkgtest [22:23:46]: testing package hipblaslt version 
7.2.4+dfsg-0ubuntu1~ppa1~26.04
autopkgtest [22:24:51]: test hipblaslt1-tests: rocm-test-launcher 
/usr/libexec/rocm/libhipblaslt1-tests/run-tests
[  PASSED  ] 3147 tests.
autopkgtest [22:25:10]: test hipblaslt1-tests: -----------------------]
autopkgtest [22:25:11]: test hipblaslt1-tests:  - - - - - - - - - - results - - 
- - - - - - - -
hipblaslt1-tests     PASS
autopkgtest [22:25:11]: @@@@@@@@@@@@@@@@@@@@ summary
hipblaslt1-tests     PASS

[ Where problems could occur ]

 * The BF16 and hipBLASLtExt API additions are new code paths that
   were not exercised in 7.1.x. A regression in these paths would manifest
   as HIPBLAS_STATUS_INTERNAL_ERROR or incorrect numerical results in
   BF16/mixed-precision GEMM on gfx90a, or as API call failures on
   gfx11XX/gfx12XX hardware using the Ext API.

 * The Sigmoid epilogue addition introduces new kernel dispatch logic.
   A regression would appear as incorrect activations or a segfault when
   HIPBLASLT_EPILOGUE_SIGMOID_EXT is requested.

 * The CPX mode fix changes the code path for complex GEMM operations.
   Although the old behaviour was a bug, any application that worked around
   the HIPBLAS_STATUS_INTERNAL_ERROR by adjusting matrix sizes may behave
   differently with the fixed path. Symptom: unexpected success (benign) or
   a different error at a different size boundary.

 * The Tensile device library (GPU kernels) is rebuilt for each upstream
   release. A regression in kernel selection heuristics could cause
   performance degradation or incorrect results for GEMM shapes that
   exercise newly added or modified kernel variants.

[ Other Info ]

 * ABI: public API is fully stable. abipkgdiff (7.1.1 vs 7.2.4) reports
   0 removed, 0 changed, 0 added public functions/variables. Internal
   symbol churn (origami:: heuristics, __hip_cuid_* build IDs) is expected
   and not ABI-relevant. SONAME: libhipblaslt.so.1.1 (unchanged).
   No debian/libhipblaslt1.symbols file exists — tracking this as a
   follow-up item.

 * This update is part of the coordinated ROCm 7.2 stack release.

 * PPA:
https://launchpad.net/~igorluppi/+archive/ubuntu/hipblaslt-7.2.4/+packages

 * Upstream version comparison:
   https://github.com/ROCm/rocm-libraries/compare/rocm-7.1.1...rocm-7.2.4

 * Target: resolute 26.04 / stonking 26.10

** Affects: hipblaslt (Ubuntu)
     Importance: Undecided
         Status: New

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2155653

Title:
  SRU: New upstream version 7.2.4

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/hipblaslt/+bug/2155653/+subscriptions


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

Reply via email to