sue-fullrun

What problem does it solve?

Manually coordinating large-scale batches of single-GPU ML training and generation jobs across heterogeneous HPC and cloud backends is error-prone, often leads to wasted GPU capacity, and lacks consistent validation, tracking, and result aggregation. This Skill eliminates those gaps by providing a strict, guardrail-heavy orchestration workflow for full scale-up experiment runs.

Core Features & Use Cases

• GPU-Occupancy-First Packing: Automatically maximizes available GPU capacity across LUMI, Snellius, Brev, AutoDL, RunPod and local backends, with hard rules against vacant GPUs both per-node and cluster-wide.
• End-to-End Run Orchestration: Manages the full workflow from preflight readiness checks, job submission, monitoring, to result summarization, with mandatory W&B and CSV ledger tracking for all runs.
• Strict Contract Enforcement: Enforces inode-safe output handling, full-path output display, dryrun-fidelity validation, and failure discipline to avoid wasted compute and broken result pipelines.
• Use Case: A researcher running 32 one-GPU image generation variants on LUMI can use this Skill to automatically pack them onto 4 full 8-GPU nodes, validate all preflight requirements, submit the jobs, monitor for GPU vacancies, and aggregate all results into a standardized final bundle synced to their local workspace.

Quick Start

Use the sue-fullrun skill to orchestrate a full batch of one-GPU ML training or generation jobs across your selected HPC or cloud backend, from preflight checks through final result aggregation.