GPU clusters

Example: a SLURM cluster with NVIDIA GPU nodes (e.g. A100 40/80 GB). Define it in your registry and submit by its --system name.

GPU billing rule

Many sites bill a fixed number of CPUs per GPU. If your registry entry sets cpus_per_gpu, labflow enforces cpus-per-task = cpus_per_gpu × --gpus so you request (and are billed) consistently.

labflow submit my_experiment --system cluster-gpu \
    --partition normal --nodes 1 --gpus 2 --time 2:00:00 --account <account>

Container

Use a CUDA-enabled Apptainer image, or your site's prebuilt GPU containers. For JAX, start from the simulation-jax template and a CUDA base image.

See HPC systems registry for the entry schema.