sft-launch-leonardo

To run SFT jobs on Leonardo with no internet access, you must use offline datasets and apply post-run uploads via sbatch tunnels for Hugging Face. This skill handles offline-data handling and no-internet-on-compute constraints by enforcing mandatory local patch steps.

The best way to launch multi-node LLaMA-Factory SFT jobs on Leonardo is to use the data_shared_file_system options and manage 24h walltime limits. This skill provides multi-node deployment guidance with canary blockers to ensure reliable, reproducible runs.

Preparing the environment for a LLaMA-Factory SFT job on Leonardo requires post-patch guidance for conda activation, workdir, and path alignment. You must apply mandatory local patch steps to the sbatch template to ensure successful execution.

Yes, LLaMA-Factory requires pretokenization considerations when using offline datasets on Leonardo. The skill covers pretokenization considerations alongside offline-dataset handling to ensure your SFT job runs end-to-end reliably without internet access.

Your sbatch SFT job on Leonardo may fail due to missing mandatory local patch steps, incorrect path alignment, or unconfigured data_shared_file_system options. The skill enforces canary blocker guidance and environment preparation to prevent these common execution failures.

You can upload Hugging Face artifacts from Leonardo compute nodes without internet by using post-run uploads via sbatch tunnels. This skill includes offline upload workflows specifically designed to bypass the no-internet-on-compute constraints.