rl-newton-cluster-ops

name: rl-newton-cluster-ops description: "Submit, monitor, sync, and ledger Moleworks Newton RL runs on Euler or Brev. Use when preparing a smoke test, launching shared-turn jobs, checking squeue/sacct/Slurm logs, syncing one run, or updating docs/experiments."

Newton Cluster Ops

Use this skill for cluster-side RL operations on the active moleworks_newton branch.

Scope

• local smoke gate before cluster launch
• choose the right submit helper
• monitor squeue, sacct, and Slurm logs
• targeted sync of one run
• keep docs/experiments accurate

Source Of Truth

Read the branch docs before improvising:

• docs/ResearcherWorkflow.md
• docs/experiments/README.md
• docs/experiments/latest.md
• docs/experiments/running.md
• docs/experiments/done.md
• docs/SHARED_TURN_DEFAULT_TRAINING.md
• cluster/README.md

For the current shared-turn dev/analytic work, prefer the Euler helpers and ledgers over older generic Daint examples in legacy docs.

Launcher Selection

• Shared-turn default recipe: cluster/submit_shared_turn_w_cabin_default.sh
• Shared-turn partially dug recovery or replay finetune: cluster/submit_shared_turn_w_cabin_partially_dug.sh
• Shared-turn no-AoA actor branch family: cluster/submit_shared_turn_w_cabin_no_aoa_actor.sh
• Generic training: cluster/submit_job.sh
• Brev only when explicitly requested: cluster/submit_job_brev.sh

Hard Rules

• Local smoke first, with W&B disabled.
• Real cluster runs use W&B.
• After submitting any real training run, verify that the run actually gets through startup before reporting it as successfully launched. sbatch success, a Brev service pid, or a stale RUNNING ledger row is not enough. Keep checking until the run either enters the training loop or fails with a concrete reason.
• Worktree launches are source-sensitive. cd into the intended worktree's cluster/ directory before submit_job.sh, submit_job_brev.sh, sync_code.sh, or sync_code_brev.sh. A shared CLUSTER_ENV_FILE is fine; do not set LOCAL_MOLEWORKS_DIR unless intentionally submitting a different tree.
• For Brev/Euler worktree smoke tests, verify the remote snapshot when code identity matters: compare sha256sum for one or two changed source files against the local worktree, then inspect the run log for the expected experiment alias and diagnostic scalars.
• Real Euler training runs should usually request JOB_TIME=24h. When composing, reviewing, or launching a real training command, set JOB_TIME=24h explicitly by default unless the user asks for a shorter run.
• Use 4h or shorter only for smoke gates, startup validation, queue/launcher probes, or intentionally bounded debugging runs. Do not use short walltimes for runs meant to judge learning quality.
• If a 24h run has not saturated at its last evaluated checkpoint, continue from the best/latest checkpoint with another long run instead of treating the partial curve as converged.
• For performance claims, prefer 24h-scale runs with pinned checkpoint benchmarks over early W&B curves from short jobs.
• Shared-turn analytic tasks keep --disable-empirical-normalization.
• Prefer Euler 1x4090. Fall back to 3090 only if the run stays pending too long.
• For Euler fee_excavation, use the normal submit_job.sh path: sync code and create a per-run snapshot. Do not set MOLEWORKS_SKIP_CODE_SYNC=1 or MOLEWORKS_SKIP_CODE_SNAPSHOT=1.
• For Euler fee_excavation on rtx_3090, cap runs at 40000 worlds. For profile sweeps, prefer a single 40000-world contract across 3090 and 4090 plus a matching RBF control; use 4090 for larger runs.
• Do not launch new real runs without updating docs/experiments/running.md.
• For exploratory reward/env sweeps, prefer diversity over duplicate seeds. Do not spend parallel GPUs on multiple seeds of the same config unless the user explicitly asks for seed robustness or the design has already narrowed to a small set of finalists. Repeated runs are not a substitute for ablations.

Smoke Gate

Prefer preparing an exact local smoke command for the user instead of running a noisy training job yourself.

Shared-turn smoke template:

export WANDB_MODE=disabled
uv run python scripts/rsl_rl/train.py \
  --task m445_excavation_shared_turn_w_cabin \
  --num-worlds 8 \
  --max-iterations 3 \
  --disable-empirical-normalization
unset WANDB_MODE

Monitoring

Prefer narrow status commands and summarize the result instead of dumping full logs into context.

ssh euler 'squeue -u $USER'
ssh euler 'sacct -j <jobid> --format=JobID,State,ExitCode,Elapsed,Start,End -P'
ssh euler 'tail -n 200 /cluster/scratch/$USER/moleworks_logs/slurm-<jobid>.out'

Post-Submit Startup Gate

For every real Euler/Brev training launch, run a startup gate before ending the turn or calling the run "launched":

1. Confirm scheduler/process state:
   • Euler: squeue immediately after submit, then sacct if the job leaves the queue.
   • Brev: confirm the service/python process is alive and tied to the expected GPU.
2. Inspect stdout/stderr for the expected startup contract:
   • expected GPU assignment (CUDA_VISIBLE_DEVICES, SLURM_JOB_GPUS, or Brev GPU list)
   • expected task, experiment alias, seed, world count, and diagnostic flags
   • expected code snapshot path for worktree launches
3. Keep watching past slow world/reset-cache construction until one of these terminal startup states is reached:
   • startup verified: W&B run id/URL exists, or the log reaches Creating PPO runner, PPO runner created, Starting training, or a first checkpoint/run directory with training artifacts.
   • startup failed: Slurm/Brev exits, stderr has an exception, OOM, reset-cache failure, invalid config, import error, or missing resource.
4. If startup fails before W&B initializes, say explicitly that there is no W&B run, capture the failing phase and stack trace headline, and update the ledger with a terminal label such as FAILED_RESOURCE, FAILED_INVALID, or FAILED_RESET_PREFILTER.
5. If the user requested a short-queue/short-walltime launch, still perform this gate. Short jobs are especially easy to lose in reset-cache or world construction before training begins.

Useful Euler probe:

ssh euler 'squeue -j <jobid> -o "%i|%T|%M|%D|%R|%b"; \
  echo ---OUT---; tail -n 220 /cluster/scratch/$USER/moleworks_logs/slurm-<jobid>.out; \
  echo ---ERR---; tail -n 160 /cluster/scratch/$USER/moleworks_logs/slurm-<jobid>.err'

If W&B is expected but no W&B URL appears yet, do not assume the run is healthy; check whether it is still constructing the world/reset cache or has failed before logger initialization.

When waiting for long training runs to reach curriculum/checkpoint gates, sleep in long intervals by default:

sleep 6000

Use shorter sleeps only for startup validation, benchmark completion, debugging failures, or when the user explicitly asks for a tighter cadence. While sleeping, avoid repeated manual polling; resume with a compact status table sourced from W&B, Slurm, checkpoint files, and benchmark markers.

If the run is pending too long, check the requested GPU type and decide whether to resubmit on 3090.

Delegated Metric And Benchmark Checks

For multi-run sweeps and any local benchmark execution, delegate noisy inspection work to sub-agents by default. Lorenzo's standing preference is that benchmark runs and benchmark-log parsing happen in a worker agent so the parent context stays clean.

Preferred sequence:

1. Ask a metrics worker to read W&B/TensorBoard/Slurm summaries first and return a compact table with latest iteration, curriculum level, success proxies, termination proxies, torque-over-limit metrics, and checkpoint availability.
2. Decide from that table which checkpoints are worth syncing and benchmarking.
3. Ask a benchmark worker to sync only selected artifacts and run local benchmarks sequentially, because the local workstation normally has one GPU. The worker must not start concurrent benchmark processes.
4. Have the worker write raw benchmark output under /tmp or the run output directory and return only report paths plus a compact result table.

Do not stream full benchmark logs into the parent conversation unless debugging a crash. The parent should keep only the decision-relevant numbers: success, close/full/partial, major terminations, tau_applied_over_limit, tau_preclip_over_limit, and any force/depth diagnostics requested by the user. If delegation tools are unavailable, run the smallest necessary local check and summarize only from generated report files.

Sync

Prefer targeted sync for one run:

cd cluster
./sync_logs.sh \
  --experiment <experiment> \
  --run-name <run_name> \
  --slurm-job <jobid>

Expected local destinations:

• logs/rsl_rl/<experiment>/<run_name>/
• logs/slurm/slurm-<jobid>.out

Treat the final folder name under logs/rsl_rl/<experiment>/ as the canonical run_name. Example:

• /cluster/scratch/$USER/moleworks_logs/rsl_rl/fee_excavation_runtime_cache_ab/2026-03-30_22-10-27_fee_excavation-prepr_runtime_cache_on_s201_4090
• canonical run_name: 2026-03-30_22-10-27_fee_excavation-prepr_runtime_cache_on_s201_4090

Use the broad ./sync_logs.sh only when the user wants a full refresh.

Ledger Rules

Treat these as mandatory for real runs:

1. Add a row to docs/experiments/running.md immediately after submit.
2. Include run_name, wandb_run, wandb_url, job id, and key artifact paths.
3. Reconcile running.md against live cluster state before reporting status.
4. When a completed pinned benchmark clearly changes the best-known checkpoint for a named condition, update docs/experiments/latest.md.
5. When a run is no longer live, preserve the latest detailed state in docs/experiments/archive/.
6. Then move the compact summary row to docs/experiments/done.md.
7. A non-live job must not remain RUNNING; mark it with one terminal label such as FAILED_INVALID, FAILED_RESOURCE, FAILED_RESET_PREFILTER, FAILED_SIGNAL, CANCELLED_PRUNED, or TIMEOUT_COMPLETE.

Reporting Format

Use one-line snapshots:

<job_id> | <run_name> | <task> | wandb_run=<...> | wandb_url=<...> | timeout=<...> | full=<...> | close=<...>

Keep the report compact and source every field from either the synced Slurm log, W&B, or the experiment ledger.