sue-nm5-env-install

What problem does it solve?

NM5 HPC sandbox compute nodes have no outbound internet access, making it impossible to install Python packages, model weights, or datasets directly on the nodes. Manually building environments on external hosts and transferring them to NM5 is error-prone, time-consuming, and risks workspace isolation breaches for scale-up ML experiments.

Core Features & Use Cases

• Automated end-to-end environment workflow: Builds conda environments on a persistent internet-connected RunPod CUDA pod, packs them into transferable tarballs, moves them to NM5, and validates them with GPU smoke tests.
• Strict workspace isolation: Ensures all environments, caches, and artifacts are scoped to the active experiment workspace, preventing cross-workspace contamination.
• Multi-stage validation gates: Runs missing inventory analysis, gap analysis debate, and pre-pack readiness checks to catch missing dependencies, weights, or datasets before transfer, eliminating failed NM5 runs.
• Use case: ML researchers running scale-up training or generation experiments on NM5 can use this skill to automatically prepare a fully functional offline environment in minutes, without manual package management or transfer troubleshooting.

Quick Start

Invoke the sue-nm5-env-install skill to automatically build your workspace's conda environment on a RunPod host, transfer it to the NM5 sandbox, and validate it passes a GPU smoke test.