palmetto-conda-scratch-guard

To manage conda environments on Palmetto HPC without violating storage quotas, create and activate heavy environments under /scratch/$USER/envs. This keeps large datasets and artifacts out of your /home directory while maintaining lightweight code and configurations locally.

To check GPU visibility and CUDA readiness before running a job on Palmetto, execute a preflight script. This script detects whether you are on a login or compute node and validates your hardware environment to prevent failed runs.

The best way to structure HPC workflows for large datasets and models on Palmetto is to enforce a storage policy. Keep lightweight code under /home and move heavy data to /scratch, using documented workflows to govern node access and environment creation.

Yes, the Palmetto scratch storage policy applies to environment caches and model artifacts. It enforces placing large artifacts, datasets, and scratch-based environments under /scratch while retaining only lightweight code and configs under /home.

Your CUDA experiments fail on Palmetto login nodes because GPUs are typically unavailable for compute tasks there. A preflight check detects login vs compute nodes and validates GPU visibility to ensure you run experiments only on appropriate hardware.