Skip to content

Conference call notes 20260325

Jump to bottom
Kenneth Hoste edited this page Mar 25, 2026 · 2 revisions

(back to Conference calls)

Notes on the 291th EasyBuild conference call, Wednesday 25 Mar 2026 (09:00 UTC / 10:00 CET)

Attendees

List of attendees (21):

  • Bibek Chapagain (Barcelona Supercomputing Centre, Spain)
  • Jorge Delgado Guerrero (University of Luxembourg)
  • Alex Domingo (Vrije Universiteit Brussel, Belgium)
  • Davide Grassano (CECAM)
  • Jasper Grimm (University of York, UK)
  • Jorge Guerra Garcia (Universidad Politécnica de Madrid, Spain)
  • Leonardo Honfi Camilo (Wageningen University, The Netherlands)
  • Kenneth Hoste (HPC-UGent)
  • Georgios Kafanas (University of Luxembourg)
  • Emik Lin (Univ. of Hong Kong)
  • Kurt Lust (UAntwerpen & LUMI User Support Team)
  • Vladislav Mogilev (Univ. of Birmingham, UK)
  • Sam Moors (Vrije Universiteit Brussel, Belgium)
  • Alan O'Cais (freelancer, HPC-UGent)
  • Jurij Pečar (EMBL, Germany)
  • Jarne Renders (Vrije Universiteit Brussel, Belgium)
  • Jan André Reuter (JSC, Germany)
  • Jörg Saßmannshausen (Imperial College London, UK)
  • Åke Sandgren (Umeå University, Sweden)
  • Alexander Wilhelmi (Goethe University Frankfurt, Germany)
  • Cintia Willemyns (Vrije Universiteit Brussel, Belgium)

Agenda

  • overview of recent developments
  • update of common toolchains
  • Q&A

Recent developments

News

Merge sprints

  • last merge sprint (Mon 16 Feb 2026)
  • future easyconfigs merge sprints planned:
    • Mon 13 Apr 2026 rescheduled to Wed 8 Apr 2026*
    • Mon 15 Jun 2026
    • Mon 17 Aug 2026
    • aiming for semi-fixed schedule every other month: 3rd Monday in even months
  • at some point a sprint to revise & close outdated open issues across different EasyBuild repos

Merged PRs

(changed made in PRs marked with * are included latest EasyBuild stable release)

  • blog/docs (merged PRs)

    • ...

  • framework (merged PRs)

    • bug fixes
      • fix handling of NVHPC toolchain with HMNS (PR #5141)
    • enhancements
      • add support for multiple modulenames in extensions (PR #4716)
      • Add ROCm toolchain based on LLVM toolchain efforts (PR #5099)
      • Allow the use of extension specific github_account so GITHUB_SOURCE template constant can be used (PR #5121)
    • changes
      • ...
    • code cleanup
      • ...
    • CI/test suite
      • ...
    • other
      • ...

  • easyblocks (merged PRs)

    • bug fixes
      • set $CIRCLECI environment variable for torchvsion to work around flaky test (that requires downloading a file from wikimedia.org) (PR #4094)
      • avoid crash in test step of custom easyblock for SuperLU if OpenMPI is not a dependency (PR #4099)
    • enhancements
      • use clean_dir instead of remove_dir on installation directory in custom easyblock for: Anaconda (PR #3934), FDTD_Solutions (PR #3933), Java (PR #3928)
      • Add ignore patterns for some RISC-V tests in LLVM easyblock (PR #4059)
      • enhance custom easyblock for Amber easyblock to support installing pmemd only (PR #4093)
    • updates
      • ...
    • changes
      • ...
    • new
      • add custom easyblock for Sentaurus (PR #4001)
    • code cleanup
      • ...
    • CI
      • ...

  • easyconfigs (merged PRs)

    • ~60 easyconfig PRs were merged since last conf call
    • bug fixes/reports
      • don't depend on ScaFaCos in LAMMPS easyconfigs when building on Arm + add patch to disable failing test on Arm due to bug in kim-api (PR #25133)
      • use NETCDF_ENABLE_DAP_REMOTE_TESTS instead of ENABLE_DAP_REMOTE_TESTS and prevent use of srun in netCDF 4.9.3 easyconfigs (PR #25490)
      • add additional patches for PyTorch 2.9.1 to fix problems with the tests + exclude inductor/test_flex* tests (PR #25492)
      • remove PyTorch (and CUDA) dependency for CellRank 2.0.2, since they're not actually needed (PR #25580)
    • enhancements
      • ...
    • (noteworthy) new software
    • noteworthy software updates
      • ...
    • cleanup
      • ...
    • changes
      • ...

Open (active) PRs

  • blog/docs (open PRs + issues)

    • ...

  • framework (open PRs + issues)

    • bugs
      • RPATH wrappers may be ignored if /tmp is mounted with noexec (issue #4910)
      • Ensure transitive dependencies for headers and libraries are always available (PR #5143)
    • enhancements
      • Proposed changes to make EasyBuild plugin-able through entrypoints (PR #4918)
        • better alternative to hooks for customizing EasyBuild
      • New experimental eb2 CLI using click (PR #4961)
      • add support for installing with bwrap (PR #5130)
    • code cleanup
      • ...
    • changes
      • Sort values for cuda_compute_capabilities templates (PR #5144)
    • CI/test suite
      • ...

  • easyblocks (open PRs + issues)

    • bug fixes/reports
      • include /usr/local Python package directory for system-level Python package installations (PR #4095)
        • fix for issue #4100: invalid command 'bdist_wheel' error when installing SYSTEM Python packages, including EasyBuild on EL10 systems
    • enhancements
      • use clean_dir instead of remove_dir on installation directory in various easyblocks
      • add sanity_pip_list parameter to verify Python package names and versions with pip list (PR #4049)
        • can be tested with --sanity-check-only on jsc-zen3
        • others can help with --sanity-check-only --include-easyblocks-from-pr 4049 --upload-test-report
        • Mikael: should definitely test CUDA-Python
        • updated to auto-enable checking of pip list when --upload-test-report, so check is required to pass for new contributions...
      • add support to install dummy python packages (PR #4063)
        • used for CUDA-wheel-bundle: PR #25266
        • use case for this is PyTorch installed with wheels: PR #25267
          • not compatible with PyTorch built from source because different compiler was used for building the wheels, but only for stuff linking to libtorch.so (rather than just using import torch)
      • enhance custom easyblocks for SCOTCH & MUMPS to add lfoss toolchain support (PR #4069)
      • Limit number of parallel compile jobs in PyTorch tests by setting $MAX_JOBS + add support for using Arm Compute Library (ACL) as dependency (PR #4096)
      • enhance custom easyblock for intel-compilers to support sysroot + dynamic linker (PR #4098)
        • see also discussion in Slack
        • first step, additional fixes needed to make sure that Intel oneAPI compilers can find their own libraries
    • updates
      • ...
    • changes
      • ...
    • code cleanup
      • ...
    • new easyblocks
      • add custom easyblock to build & install ROCm components (PR #3861)
        • see PR for ROCm 6.4.1 easyconfigs: PR #25576
      • custom easyblock for ESPResSo (PR #4066)
      • add custom easyblock for DeepSpeed (PR #4082)

  • easyconfigs (open PRs + issues)

    • bug fixes/reports
      • ...
    • enhancements
      • ...
    • (noteworthy) new software
      • ...
    • software updates
      • PyTorch 2.9.1 w/ CUDA (multiple toolchains)
        • PyTorch performance benchmarking results by Alex: issue #931
    • changes
      • bump OpenBLAS to v0.3.32 as dependency for FlexiBLAS 3.5.0 w/ GCC 15.2.0 (PR #25603)

Common toolchains

  • {foss/intel/lfoss}/2026.1
    • revised versioning scheme for common toolchains, moving away from <year>{a,b}
      • aiming to be closer to one common toolchain version per year (instead of two)
    • likely to be included with next EasyBuild release...
    • relevant PRs tagged with 2026.1 label
  • 2026.1 toolchain definitions are WIP
    • hopefully to be included in upcoming EasyBuild release...
    • foss/2026.1 component versions (all latest release)
    • intel/2026.01 component versions
      • GCC 15.2
      • intel-compilers 2025.3.2
      • impi 2021.17.2
      • imkl 2025.3.1
    • lfoss/2026.01 component versions
      • like foss/2026.01, but with LLVM 21.1.8
      • need to fix use of -march=native on Arm with LLVM, should use -mcpu=native instead
      • OpenBLAS 0.3.31 is using grace as target, trouble when -march=native is used, which leads to missing some CPU features and trigger build failure; fixed when using -mcpu=native
  • probably worth waiting for CUDA 13.3
    • to avoid regression with C++20 in older CUDA 13.x

Q&A / other

  • Jörg: incorrect OpenMPI version in ORCA
  • Alan: AI-assisted contribution to framework
  • Jörg: support for Intel Sapphire Rapids?
    • should work well, already one of the CPU targets in EESSI
    • builds one Sapphire Rapids can be quite slow, unclear why, may be because of older OS versions being used

Clone this wiki locally
⚡