nemo-rl-brev-etiquette

작성자: nvidia

/home/ubuntu/RL에서 제한된 작업 공간 디스크, 더 큰 /ephemeral 볼륨, 그리고 선택적…을 사용하는 NeMo-RL 에이전트를 위한 Brev 인스턴스 운영 가이드

npx skills add https://github.com/nvidia/skills --skill nemo-rl-brev-etiquette

Brev Etiquette

Operate as though /home/ubuntu/RL is the source checkout and /ephemeral is the working storage for generated experiment state. Keep the repo small, reproducible, and easy to inspect. Move bulky run outputs to /ephemeral before launching anything expensive.

Storage Rules

  • Keep code edits, small config changes, committed experiment hypotheses, and concise reproducibility records under /home/ubuntu/RL.
  • Put generated experiment assets under /ephemeral, including checkpoints, run logs, Ray temp directories, W&B offline files, profiler traces, evaluation dumps, rollout samples, and per-experiment artifacts.
  • Keep reusable caches under one shared /ephemeral cache root per user, not under each experiment. This includes Hugging Face models, dataset caches, PyTorch caches, Triton caches, uv caches, and pip caches.
  • Before a campaign or long run, check capacity with df -h /home/ubuntu/RL /ephemeral and avoid starting if /ephemeral is missing or nearly full.
  • Create a campaign root such as /ephemeral/nemo-rl/${USER:-ubuntu}/nemo-rl-auto-research/<campaign> and use one subdirectory per experiment.
  • Do not leave large files, cache directories, or generated outputs in the git checkout. If a tool defaults to the repo, override its output/cache path before running it.

Environment Secrets

  • Treat /home/ubuntu/RL/.env as the local secret store. It may contain keys such as WANDB_API_KEY, HF_TOKEN, or HUGGING_FACE_HUB_TOKEN.
  • Before any run that may need external auth, load /home/ubuntu/RL/.env when it exists. Never print, cat, log, commit, or summarize secret values.
  • If /home/ubuntu/RL/.env is absent, or a required key is still unset after loading it, remind the user to add the needed key to that file before launching authenticated work.
if [ -f /home/ubuntu/RL/.env ]; then
  set -a
  . /home/ubuntu/RL/.env
  set +a
else
  echo "Missing /home/ubuntu/RL/.env; add required keys such as WANDB_API_KEY or HF_TOKEN before authenticated runs."
fi

Auto-Research Pattern

When using nemo-rl-auto-research, keep the git ledger in the repo and heavy evidence on /ephemeral.

if [ -f /home/ubuntu/RL/.env ]; then
  set -a
  . /home/ubuntu/RL/.env
  set +a
fi

BREV_ROOT=/ephemeral/nemo-rl/${USER:-ubuntu}
CACHE_ROOT=$BREV_ROOT/cache
CAMPAIGN_ROOT=$BREV_ROOT/nemo-rl-auto-research/<campaign>
EXP_DIR=$CAMPAIGN_ROOT/<experiment>
mkdir -p "$EXP_DIR"/{logs,checkpoints,artifacts,ray,tmp,wandb}
mkdir -p "$CACHE_ROOT"/{huggingface,torch,triton,uv,pip,xdg,wandb}

export HF_HOME=$CACHE_ROOT/huggingface
export HF_HUB_CACHE=$HF_HOME/hub
export HF_DATASETS_CACHE=$HF_HOME/datasets
export TRANSFORMERS_CACHE=$HF_HOME/transformers
export TORCH_HOME=$CACHE_ROOT/torch
export TRITON_CACHE_DIR=$CACHE_ROOT/triton
export UV_CACHE_DIR=$CACHE_ROOT/uv
export PIP_CACHE_DIR=$CACHE_ROOT/pip
export XDG_CACHE_HOME=$CACHE_ROOT/xdg
export WANDB_CACHE_DIR=$CACHE_ROOT/wandb
export RAY_TMPDIR=$EXP_DIR/ray
export TMPDIR=$EXP_DIR/tmp
export WANDB_DIR=$EXP_DIR/wandb

Record the absolute /ephemeral paths in the nemo-rl-auto-research TSV fields for log path, checkpoint path, artifacts, shared cache root, and command. If the TSV itself may grow large, store the full TSV in /ephemeral and keep a small pointer file or summary in the repo.

Launch Checklist

  • Inspect disk first: df -h /home/ubuntu/RL /ephemeral.
  • Choose a unique /ephemeral run root before editing recipes or launching jobs.
  • Reuse a shared cache root such as /ephemeral/nemo-rl/${USER:-ubuntu}/cache across experiments unless a run explicitly requires a clean cache.
  • Override recipe output paths, logger paths, checkpoint paths, and temp paths to point under the experiment directory.
  • Override cache paths to point under the shared cache root.
  • Stream stdout/stderr to $EXP_DIR/logs/run.log or an equivalent file under /ephemeral.
  • Periodically check disk during long runs with df -h /ephemeral and stop gracefully if the volume is approaching exhaustion.
  • At the end, summarize the important metrics and paths in the repo ledger; do not copy bulky artifacts back into /home/ubuntu/RL.

Cleanup

  • Clean only files that belong to the current campaign or experiment.
  • Prefer pruning clearly named experiment directories under /ephemeral/nemo-rl/...; never remove shared caches or another user's run directory without an explicit instruction.
  • Preserve enough small metadata in the repo to reproduce a result after /ephemeral is cleaned.

nvidia의 다른 스킬

compileiq-debug
nvidia
무언가 잘못되었을 때 사용: Search()가 멈추거나, 모든 평가가 INVALID_SCORE를 반환하거나, 점수가 개선되지 않거나, 모든 설정이 동일한 숫자를 반환하거나, ptxas 오류 등이 발생할 때
create-github-pr
nvidia
gh CLI를 사용하여 GitHub 풀 리퀘스트를 생성합니다. 사용자가 새 PR을 만들거나, 코드 리뷰를 제출하거나, 풀 리퀘스트를 열고자 할 때 사용합니다. 트리거 키워드 -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
다른 열린 이슈들을 스캔하여 주어진 PR이 함께 수정하거나 실수로 망가뜨릴 수 있는 이슈를 찾습니다. 인접 수정 기회와 모순 위험을 file:line…과 함께 출력합니다.
fhir-basics
nvidia
에이전트에게 FHIR R4 API의 작동 방식, 사용 가능한 리소스, 검색 매개변수를 사용한 쿼리 방법, 모든 응답 형식을 올바르게 파싱하는 방법을 가르칩니다…
compileiq-validate-result
nvidia
검색이 완료된 후, 속도 향상을 청구하거나 ACF를 발송하기 전에 사용합니다. dump_results CSV를 로드하고, 상위 K개 후보(단일 목표)를 추출합니다…
changelog-audit
nvidia
릴리스 전에 Warp CHANGELOG.md를 감사합니다: 누락된 항목 복구, 사용자 영향별 정렬, 항목 언어 다듬기, 줄 바꿈, (릴리스 브랜치 모드) 비교 업데이트…
maintain-dynamic-plugins
nvidia
NeMo Relay 동적 플러그인 로더, 매니페스트, Rust 네이티브 SDK, gRPC 워커 프로토콜, Python 워커 SDK, 문서, 테스트 및 릴리스 워크플로 커버리지를 유지 관리합니다.
dgx-diagnose
nvidia
일반적인 DGX Station GB300 문제 진단 — CUDA 충돌, 잘못된 GPU 타겟팅, vLLM/SGLang 컨테이너 버그, MIG 상태 문제, NVLink/Fabric Manager 오류,…