brev-etiquette

What problem does it solve?

Brev instance runs can fail or become painful when storage is limited in the repo checkout while experiment outputs, caches, logs, and checkpoints grow during long NeMo-RL jobs.

Core Features & Use Cases

• Storage discipline for reproducibility: Treat /home/ubuntu/RL as the source checkout and /ephemeral as the working storage for all generated experiment state.
• Shared cache strategy: Centralize Hugging Face, PyTorch, Triton, uv, pip, XDG, and W&B caches under a shared /ephemeral cache root to reduce recompute and avoid filling the checkout.
• Secret handling for authenticated workflows: Load /home/ubuntu/RL/.env when present, and never print or log secret values while enabling token-based downloads and integrations.
• Operational campaign layout: Use a predictable /ephemeral/nemo-rl/.../auto-research/<campaign>/<experiment> structure with per-experiment logs, checkpoints, artifacts, Ray temp, and W&B offline state.

Quick Start

Run your Brev NeMo-RL campaign by exporting all relevant cache and temp environment variables to /ephemeral and keeping generated outputs under the experiment directory while leaving only small, reproducible records in /home/ubuntu/RL.