[bartgol]

The machine is now called ghci instead of ghci-snl, and some settings have changed (e.g., BLAS libs location).

[BFB]

---

The new images should come online tomorrow, so this PR should be merged then.

Notice that the string -snl has been removed from machine names as well as from runner names. The reason is that we should be able to run on the e3sm-project/containers images, REGARDLESS of the host machine we are on. Yes, the snl-hosted runners do add some extra layers on top of the e3sm images, but only to set proxies and certificates, they do NOT change the software stack or the relevant env stuff. If we later install e3sm runners on another machine, the workflow files SHOULD be able to run without any modification (provided that those runners have the e3sm label, as well as the relevant compiler label.

[bartgol]

I killed CI, since there is no point running until the self-hosted runners run on the new images.

[Copilot]

Copilot review overview

🟡 Changes recommended

Stale machine registrations and hard-coded host architectures undermine the rename and host-independent runner objective.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Update C-ACTS registry for renamed machines · New
• Use the official GitHub spelling · New

What changed in this PR

Renames GHCI machine configurations, updates image settings, and migrates CI jobs to generic e3sm runners.

Changes:

• Renames GHCI machine definitions and baseline paths.
• Updates BLAS, Kokkos, and architecture settings.
• Migrates EAMxx workflows to new machine and runner names.

File	Description
components/​eamxx/​scripts/​machines_specs.py	Renames standalone GHCI machines.
components/​eamxx/​cmake/​machine-files/​ghci.cmake	Adds shared GHCI configuration.
components/​eamxx/​cmake/​machine-files/​ghci-intel.cmake	Updates Intel machine name and include.
components/​eamxx/​cmake/​machine-files/​ghci-gnu.cmake	Updates GNU paths and BLAS configuration.
components/​eamxx/​cmake/​machine-files/​ghci-cuda.cmake	Updates CUDA naming and BLAS configuration.
cime_config/​machines/​config_machines.xml	Renames the CIME machine and paths.
cime_config/​machines/​cmake_macros/​ghci.cmake	Disables deprecated Kokkos 4 code.
cime_config/​machines/​cmake_macros/​ghci_intel.cmake	Adds Intel Kokkos and MKL settings.
cime_config/​machines/​cmake_macros/​ghci_gnugpu.cmake	Adds CUDA Kokkos settings.
cime_config/​machines/​cmake_macros/​ghci_gnu.cmake	Adds GNU OpenMP settings.
.github/​workflows/​eamxx-v1-testing.yml	Migrates CIME tests to generic runners.
.github/​workflows/​eamxx-scripts-tests.yml	Updates script-test runner and machine.
.github/​workflows/​eamxx-sa-testing.yml	Updates standalone test runners and machines.
.github/​workflows/​eamxx-sa-sanitizer.yml	Updates sanitizer runner configuration.
.github/​workflows/​eamxx-sa-coverage.yml	Updates coverage runner configuration.
.github/​workflows/​dexpr-testing.yml	Updates runner documentation.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[bartgol]

Considerations on CI fails/passes:

• we changed the BLAS impl, from Netlib-LAPACK to OpenBlas for gnu and gnugpu/cuda builds. For intel, we are still using MKL libs, just like before. This means that
  • eamxx-sa:
    • gnu/cuda: expected to PASS in single-precision (sp), and FAIL in double-precision (dbg, opt) since we only use BLAS in homme, which is not built in sp. More specifically, we call BLAS routines in the dyn<->pg2 remap layer, which in standalone tests is only used for two tests, which end up being the ones failing. The fpe build is also in double-precision, but has NO baseline cmp tests.
    • intel: same libs as before (MKL), so intel-openmp / opt is expected to pass.
  • eamxx-v1:
    • intel: what i said above about BLAS would suggest it should PASS. But this case has been failing for a long time, due to OMPI issues, so the baselines are likely VERY old. Expected to fail.
    • gnu/gnugpu: expected to fail across the board, for the same reason why the 2 pg2 tests fail in standalone: we changed BLAS lib implementation, resulting in non-bfb changes. Only one not expected to fail is the ne4_ne4 test (no pg2 remap there).

[mahf708]

woohoo 🎉