rl-standard-launch-leonardo

To launch offline SkyRL experiments on CINECA Leonardo, use sbatch wrappers under hpc/skyrl_yaml/leonardo with a configured uv venv and writable apptainer sandbox. This avoids the traditional hpc.launch workflow for single-node and multi-node runs.

Yes, you can run multi-node reinforcement learning jobs on Leonardo without internet access. The Skill provides guardrails for no-internet environments and uses offline data staging to support GSM8K, MATH/aime, and OPD workflows.

The best way to start a GSM8K canary job on Leonardo is to run sbatch sbatch_gsm8k_canary.sh from the leonardo launcher directory. This tests your offline SkyRL setup using the sbatch wrappers and uv venv.

Yes, the SkyRL launcher supports OPD and distillation workflows on a multinode grid. It handles single-node and multi-node RL runs using sbatch wrappers and offline data staging on CINECA Leonardo.

Yes, you need a writable apptainer sandbox to run MATH/aime experiments on Leonardo. The sandboxed execution environment is required alongside sbatch wrappers, offline data staging, and a configured uv venv.

You should avoid the traditional hpc.launch workflow for offline SkyRL experiments because it lacks the necessary guardrails for no-internet environments. The Skill uses sbatch wrappers and a writable apptainer sandbox to manage offline data staging directly.