Check CUDA compute capabilities in an easystack plan¶
What it does: stack cuda-cc reads one or more easystacks as a plan,
closes their dependencies on the robot tree, keeps the CUDA-dependent
easyconfigs, and checks each cuda_compute_capabilities list against:
the CUDA toolkit version that recipe depends on;
EESSI accelerator target requirements (for example zen4 and grace);
known easyblock map limits (for example LAMMPS in software-layer #1461).
Why you’d use it: this catches the PR-1461 failure class before the bot
runs, where capability requests such as 9.0a or 12.0f do not match
the selected CUDA release or target matrix.
eb-stack stack cuda-cc \
--easystack tests/cuda_cc_fixtures/easystacks/pr-1461-failing.yml \
--easyconfigs tests/cuda_cc_fixtures/easyconfigs
The command prints:
LAMMPS-2025.1-foss-2026.1-CUDA-12.6.0.eb: CUDA 12.6 drops compute capabilities 12.0f, 9.0a (software-layer #1461)
LAMMPS-2025.1-foss-2026.1-CUDA-12.6.0.eb: CUDA 12.6 is below grace-cc90 minimum CUDA 13.1 (software-layer #1461)
LAMMPS-2025.1-foss-2026.1-CUDA-12.6.0.eb: LAMMPS easyblock map rejects compute capabilities 12.0f, 9.0a: LAMMPS easyblock has no mapping for these compute capabilities (software-layer #1461)
LAMMPS-2025.1-foss-2026.1-CUDA-12.6.0.eb: missing target capabilities 9.0 for grace-cc90 (aarch64/nvidia/grace) (software-layer #1461)
LAMMPS-2025.1-foss-2026.1-CUDA-12.6.0.eb: missing target capabilities 9.0 for zen4-cc90 (x86_64/amd/zen4) (software-layer #1461)
checked=1 findings=5 targets=2
Each line names the easyconfig file and one incompatible combination.
checked= is how many CUDA-dependent recipes had an explicit
cuda_compute_capabilities list. findings= is how many failures were
found. Any finding makes the command exit 1.
Use --target to check a subset of targets. Use --cuda-table,
--accelerator-targets, and --easyblock-map to replace the bundled
tables.