Monitor a run

Runbook

On launch

1. Find the run dir and read the resolved configs at {run_dir}/configs/latest/resolved/ (start with rl.json, or orchestrator.json on local runs). Read the launch command from {run_dir}/configs/latest/command.txt. The launch TOML is copied verbatim to {run_dir}/configs/latest/rl.toml. The run dir is {output_dir}/{run_name} — run.name auto-generates as <envs>--<model>--<short-id>, so if you only know the output dir, pick the most recently modified subdirectory (ls -t {output_dir} | head -1) or read run.name from the launch command.
2. Confirm all processes are alive and the run is making progress.
3. Write the initial summary into {run_dir}/STATUS.md.

Recurring check-ins

Default cadence: 1 hour (researcher can override). At each check-in:

1. Confirm processes are alive.
2. Grep logs for errors/warnings; note current step and key metrics.
3. Append an entry to {run_dir}/STATUS.md (never overwrite):

## YYYY-MM-DD HH:MM UTC

**Step**: {current_step} / {max_steps}
**Health**: {Healthy | Degraded | Down}

**Progress**: reward/mean, seq_len, truncation, eval scores, env-specific metrics.
**Stability**: entropy, mismatch_kl, grad_norm — flag spikes.
**Performance**: trainer vs orchestrator step time, env lag, inference pressure.

**Notes**: anything unusual (errors, restarts, hangs). Omit if nothing notable.

In W&B, each project auto-gets an "overview" saved view (train / eval / stability / performance sections) on its first run — use it for a quick check instead of the auto-generated default workspace.

Restarting a run

Never restart unless the researcher explicitly asked. Confirm the exact restart command and the conditions that warrant one.

Never run kill or launch commands yourself. Hand the researcher the exact command and let them run it; after a restart, verify all processes are back up and progress resumed before the next check-in.

Reference

Where to find things

• {run_dir}/configs/latest/ — the current attempt's command, launch TOML, and resolved/ JSON files. Each launch stays under configs/attempt_<n>/.
• {run_dir}/logs/latest/ — the current attempt's logs (each launch gets logs/attempt_<n>/; resumes never overwrite earlier attempts). See below.
• {run_dir}/monitors/file/ — the metrics, and the traces with the annotations about them (see Episodes below).

Dashboard

uv run dashboard [output_dir ...] (default outputs/, or $PRL_OUTPUT_DIR if set; several dirs can be tracked at once) serves a local web dashboard at http://localhost:7788 with four views per run: metrics (the W&B overview sections, read from metrics.jsonl), per-attempt config files, a rollout trace viewer with per-token overlays (advantage, trainer logprob, entropy, KL mismatch, stable/loss/content masks), and merged component logs. It only reads the run dirs — safe to run against a live run. --port/--host pick the bind address; a taken port automatically bumps to the next free one, so several dashboards run side by side without coordination. GPU deps live behind the gpu extra, so uv sync --extra dashboard && uv run dashboard works without the training stack (e.g. on a head node).

Daemon (auto-start): launchers auto-start one dashboard per host per user and a live one absorbs each new run's output dir automatically — see the dashboard skill for discovery, kill/restart commands, and --isolated. The short version: the live port can differ from 7788 (a taken port bumps), so read the discovery file:

cat ~/.cache/prime-rl/dashboard/daemon.json   # {"pid": ..., "url": "http://localhost:<actual port>"}
ps aux | grep PRL::Dashboard                  # the daemon's process title

Verify liveness with curl -sf <url>/api/runs and hand the researcher the url.

Logs

{run_dir}/logs/latest/
├── trainer.log                # rank 0 stdout
├── orchestrator.log           # orchestrator stdout
├── evals.log                  # SFT online-eval evals stdout
├── inference.log              # vLLM stdout
├── trainer/
│   ├── node_*.log             # per-node (multi-node only)
│   └── torchrun/              # per-rank stdout/stderr
├── inference/
│   ├── node_*.log             # per-node (multi-node only)
│   └── router.log             # the single global router (multi-node only; single-node logs it in inference.log)
└── envs/{train,eval}/{env_name}.log    # one log file per env

SLURM batch logs are under {run_dir}/launcher/logs/*job_*.log.

Usually tailing trainer.log, orchestrator.log, and inference.log is enough. Drop into per-node or per-rank logs only when debugging. All logs are loguru with HH:mm:ss LEVEL message; levels: DEBUG, INFO, SUCCESS, WARNING, ERROR.

Scan for problems:

grep -E "WARNING|ERROR" {run_dir}/logs/latest/{trainer,orchestrator,evals,inference}.log
grep -E "WARNING|ERROR" {run_dir}/logs/latest/envs/{train,eval}/*.log

Metrics

All metrics print to the console log (and W&B when configured).

Progress — orchestrator log. Rollout metrics mirror the episode/trace hierarchy, at two levels:

• {scope}/{subset}/<metric>/<stat> — episode-level facts only: the token/turn/branch counts, summed over an episode's traces.
• {scope}/{subset}/<agent>/<metric>/<stat> — every trace-level metric (reward, truncation, errors, timing, env metrics, curriculum admission, eval scores), keyed by agent name so seats never mix. Flat over that agent's traces: one sample is one trace, so an in-episode fan-out like n solvers contributes n samples.

scope is train/agg (all train envs) or train/<env> (eval/<env> for eval); subset is all (every rollout) or effective (admitted, clean, and trainable). Single-agent envs have one agent — usually agent — and one trace per episode, so both levels agree; multi-agent envs name each seat (proposer, solver, judge, …).

Stability — trainer log:

Performance — trainer and orchestrator step independently, so comparing step times shows who's waiting on whom.

The trainer warns when batch wait time exceeds active trainer time. Add inference nodes when this warning persists. The orchestrator warns when policy wait time exceeds active orchestrator time. Add trainer nodes when this warning persists. The orchestrator also warns when it discards more than half of an episode window and reports stale, errored, and no-signal counts.

orchestrator.constant_trainer_batch_size defaults to true. It keeps each rollout batch at orchestrator.batch_size effective episodes. Set it to false for faster collection with variable trainer batch sizes.

For live vLLM stats, query Prometheus directly:

curl -s http://localhost:8100/metrics | grep -E "num_requests|gpu_cache_usage"  # engine port (8000 is the router)
# vllm:num_requests_running, vllm:num_requests_waiting, vllm:gpu_cache_usage_perc (→1.0 = KV cache saturated)

Episodes

{run_dir}/monitors/file/metrics.jsonl                              # every metric row, tagged by producer
{run_dir}/monitors/file/traces/stream/00000.jsonl.zst              # every episode, appended as it arrives: sealed chunks ...
{run_dir}/monitors/file/traces/stream/00001.jsonl                  # ... and the live one, plain text
{run_dir}/monitors/file/traces/stream.index.jsonl                  # one compact row per episode, with its chunk and byte offset
{run_dir}/monitors/file/traces/annotations/{producer}/00000.jsonl  # trace updates: orch ship-time facts, trainer per-token streams
{run_dir}/monitors/file/traces/annotations/{producer}.index.jsonl  # each update's scalars and where its record sits

Everything the file monitor dumps lives under monitors/file/; nothing is written there when the monitor is off. The traces and everything written about them sit under traces/. Each stream is a directory of numbered chunks — the writer rolls to a new chunk at monitors.file.chunk_bytes (5 GiB) and, with monitors.file.compress (on), seals the full one with zstd in the background; a finished run seals its live chunk too. Sealed chunks use seekable frames, so a seek still costs one frame, and zstd -dcf streams them together with the plain live chunk. Every index is named for the stream it indexes and sits beside it. Those indexes are what keep reading a run cheap: a consumer browses them instead of the streams, and seeks by the chunk and offset they carry to read a single episode or its token streams. Both are derived, so deleting them only costs a reader the work of rebuilding what it needs. The stream holds native vf.Episode records (training tensors excluded; per-token floats rounded to monitors.file.float_decimals, 4 by default), one line per episode in arrival order, whatever kind of work it did — including trace-less failures, curriculum-rejected work, and work that never enters a batch, so it is crash-durable. Each record carries its provenance: env (id plus the orchestrator's name), full task, group (id), and run.

A trace has several steps, so each is stamped as its own event rather than implied by where the record sits. The file monitor stamps info.kind and info.dispatch/ info.arrival ({step, time} each) as an episode lands; the ship-time annotation adds info.effective and info.ship — the orchestrator step whose batch shipped the cohort, or for eval the step that produced the policy it measured. Staleness is ship.step - dispatch.step. Only effective ties to a step; all is the whole stream.

Everything learned after arrival is an append-only trace update keyed by trace_id, one file per producer so each has a single writer: the orchestrator records cohort membership, the scalar advantage and per-branch advantage streams; the trainer records its recomputed per-token logprobs and entropies. Readers fold the updates onto the stream records, newest winning.

wc -l {run_dir}/monitors/file/traces/stream.index.jsonl
zstd -dcf {run_dir}/monitors/file/traces/stream/* | jq '.traces[].rewards'
zstd -dcf {run_dir}/monitors/file/traces/stream/* | jq 'select(.ok | not) | {id, env: .env.id, errors}'
jq '{trace_id, info}' {run_dir}/monitors/file/traces/annotations/orch.index.jsonl

The batches consumed by the trainer are shipped over ZMQ by default, so nothing binary is written. With rollout_transport.type = "filesystem" they land at {run_dir}/batches/step_{n}/rank_<rank>.bin (one packed micro-batch file per trainer DP rank).

Common failure modes

A few warnings are normal. Escalate when errors are persistent, growing, or hit a large fraction of rollouts.

• Env workers: exceptions in env code, timeouts, sandbox errors, OOM kills (most common source — runs user code).
• Orchestrator: empty/errored rollout spikes, weight-broadcast failures, checkpoint errors.
• Trainer: NCCL/CUDA errors, OOM, NaN loss or gradients.
• Inference: NCCL/CUDA errors, OOM, request timeouts.

Process tree

All processes use setproctitle so they're visible in ps/htop/pstree:

PRL::Launcher
├── PRL::Inference          (vLLM server, GPU 0)
├── PRL::EnvServer          (verifiers' ZMQ env server, run in-process; one per train/eval source)
│   └── Verifiers::EnvWorker0..N
├── PRL::Orchestrator       (CPU-only; connects to each env server)
├── torchrun
│   └── PRL::Trainer        (GPU 1+)
└── tail trainer.log

For multi-node runs, trainer and inference processes are on separate nodes — use srun or ssh to inspect them.