ARM (aarch64) clusters¶
Example: an ARM SLURM cluster (e.g. Fujitsu A64FX nodes). You describe the system
once in your registry (~/.claude/contexts/hpc-systems.yaml); labflow renders the
SLURM script from that entry. The --system name is whatever you call it there.
Submit¶
labflow submit my_experiment --system cluster-arm --partition normal \
--nodes 4 --time 2:00:00 --account <account>
ARM-specific notes¶
- Compile native. Don't rely on x86 wheels — build from source for
aarch64, or use an Apptainer image compiled for the target CPU (e.g. GCC-mcpu=<cpu>). - OpenMP placement. Set
OMP_PROC_BIND=TRUEandOMP_NUM_THREADS=${SLURM_CPUS_PER_TASK}for thread-scaling performance. - Run from scratch. Jobs should run from the parallel filesystem (your registry's
scratch:path), not$HOME.
See HPC systems registry for the entry schema.