rl-standard-job-cleanup

name

rl-standard-job-cleanup

description

Preserve + publish a finished STANDARD (non-agentic GRPO) SkyRL RL checkpoint — the Delphi/rlvr/dapo math-and-reasoning cells launched via rl-standard-launch-leonardo (raw sbatch of hpc/skyrl_standard/leonardo/*, logger=console, NO Harbor/Daytona/trace_jobs). Covers: cancel pending retries, pick the BEST checkpoint by the trailing-5 EMA of reward via `parse_skyrl_metrics.py --format standard` (chain-aware, capped at the latest saved step), flatten weights to repo root, secret-scan, `hf upload` (Leonardo sbatch-tunnel) to laion/<run>-<step>-<size>B with the size suffix DERIVED FROM THE EXPORTED WEIGHTS (never the base-model name), DB register (--training-type RL + cross-user FK pre-check) ONLY for DB-registerable cells, clean up, and fire the Delphi downstream eval suite on the post-RL ckpt (defers to eval-standard-launch §5b). The ONLY artifacts are the model + the metric CSVs/report + the tracker scores — there is NO trace dataset. Use when a standard/non-agentic GRPO RL run finishes and needs uploading + (maybe) registering. DISTINCT from rl-agentic-job-cleanup, which is the AGENTIC (Harbor/Daytona) path with a companion trace dataset.

⚠ Do not add comments to YAMLs. Report your recommendations directly to the supervisor.

rl-standard-job-cleanup

Cleanup for a finished standard (non-agentic) GRPO SkyRL run on Leonardo. It publishes laion/<run_name>-<step>-<size>B with root-level weights and metric CSVs/report; add a Supabase models row only for DB-registerable series (§7).

Not the agentic cleanup. Standard GRPO has no trace_jobs/, trainer_log.jsonl, trial, or batch-error outputs.

What this cleanup drops vs the agentic one

• No training-trace dataset (make_and_upload_trace_dataset / penfever/<job>) — no Harbor rollouts, no trace_jobs/. Omit the agentic "Training Traces" README section.
• No trial_stats.csv / batch_errors — per-trial result.json is agentic only.
• No trainer_log.jsonl — logger=console; the .out chain is canonical. Publish metrics.csv, vllm_metrics.csv, report.md, and reward_plot.png.

Cross-cutting rules

• hf upload, NEVER hf upload-large-folder (deprecated stub + deadlocks on HF LFS 429s). hf upload is additive (safe to re-run; resumes from .cache/huggingface/). Never use huggingface_hub.upload_folder() without delete_patterns=[] — it clobbers files absent locally.
• --private is a no-value flag — default is PUBLIC; omit it.
• Run parse_skyrl_metrics.py from the otagent conda env (matplotlib/pandas; the RL uv venv lacks them).
• Leonardo login-node killer (~100 s): a login-node hf upload is SIGKILLed at ~100 s regardless of detach. Use the sbatch+tunnel pattern (.agents/ops/leonardo/ops.md → "Leonardo HF Upload — Use sbatch, NOT the Login Node"). The tunnel depends on the ~/.ssh/leonardo_daytona step-ca cert (~12 h) — a fast publickey failure means refresh the cert.
• GPFS hygiene: never find/du on $WORK; locate the .out via scontrol show job <id> -o (StdOut=/%Z) or a depth-1 ls; rm -rf cleanup runs detached (no sizing first).
• Run on a login node (login02/03/04 — login01 false-drains).

0. Cancel pending retries

squeue -u $USER --format='%.18i %.90j %.8T' | grep <RUN_NAME>
scancel <retry_job_ids>

1. Locate the run dir + the FULL .out chain

Standard RL writes to $WORK/rl_ckpts/<RUN_NAME> ($WORK=/leonardo_work/AIFAC_5C0_290/bfeuer00):

WORK=/leonardo_work/AIFAC_5C0_290/bfeuer00
RUN_DIR=$WORK/rl_ckpts/<RUN_NAME>
ls -d $RUN_DIR/exports/global_step_*/        # HF-exportable checkpoints (every hf_save_interval steps)
cat $RUN_DIR/latest_ckpt_global_step.txt     # last saved step — the selection CAP

The .out files are %x_%j.out in the sbatch CWD — hpc/skyrl_standard/leonardo/<JobName>_<jobid>.out (standard RL has no experiments/<job>/logs/ dir). A chain restart = one .out per link:

cd $WORK/code/OpenThoughts-Agent/hpc/skyrl_standard/leonardo
ls -lt *<RUN_NAME>*_*.out      # collect ALL chain links (sorted by jobid/time)

2. Pick the BEST checkpoint + emit the metrics — parse_skyrl_metrics.py --format standard

Computes trailing-5 reward/avg_raw_reward EMA (α=1/3), intersects on-disk exports, caps at the latest saved step, and emits metrics.csv, vllm_metrics.csv, report.md, and reward_plot.png. Run from otagent.

Tool-presence pre-check: if --format standard is missing, sync only this file, never a full git pull:

cd $WORK/code/OpenThoughts-Agent
scripts/analysis/parse_skyrl_metrics.py --help 2>/dev/null | grep -q 'standard' || \
  grep -q "format" scripts/analysis/parse_skyrl_metrics.py || echo "MISSING --format standard"
# If missing, fetch JUST the tool (does NOT touch other live-applied edits):
git fetch origin penfever/working
git checkout origin/penfever/working -- scripts/analysis/parse_skyrl_metrics.py

The CLI takes one log and output folder, not individual .out files. Stage the real chain links in a clean folder, excluding *_rayw*.out and *_rayhead*.out:

OUTLOGS=$WORK/rl_cleanup/<RUN_NAME>/outlogs
OUT=$WORK/rl_cleanup/<RUN_NAME>/metrics
mkdir -p $OUTLOGS
# Copy ALL real chain-link TRAINING .out logs (one per restart link) — NOT the _rayw*/_rayhead worker logs:
cd $WORK/code/OpenThoughts-Agent/hpc/skyrl_standard/leonardo
for f in *<RUN_NAME>*_*.out; do
  case "$f" in *_rayw*|*_rayhead*) continue;; esac
  cp "$f" $OUTLOGS/
done
ls $OUTLOGS/

/leonardo_work/AIFAC_5C0_290/bfeuer00/miniforge3/envs/otagent/bin/python \
  scripts/analysis/parse_skyrl_metrics.py \
  --format standard \
  --run_dir $RUN_DIR \
  --save_every 20 \
  $OUTLOGS \
  $OUT

• --save_every MUST equal the run's trainer.hf_save_interval (Delphi default 20). Verify in the .out launch line (trainer.hf_save_interval=…). Selection excludes the first save and starts at 2*save_every.
• CHAIN-AWARE — stage ALL real .out links, not one. A chain's first link(s) often have 0 train lines (engine bring-up); a single mid-chain .out under-covers the EMA window and can pick the wrong step. First-seen-wins per step, so overlapping links are safe to include.
• The tool prints the EMA table + CHOSEN STEP: <best> and writes it into report.md. Record BEST=<best>; the export to publish is $RUN_DIR/exports/global_step_<BEST>/policy/.
• NO STEP CHOSEN (e.g. only global_step_20 exists, or no reward lines parsed): inspect report.md's EMA table; a scancelled-early run's only eligible ckpt may be the largest saved multiple of save_every ≤ cap — fall back to that and note it. Never publish the first save (global_step_20).

3. Flatten model files to the upload-dir ROOT

HF model files MUST sit at the base of the uploaded dir — not under policy/:

UPLOAD_DIR=$WORK/rl_cleanup/<RUN_NAME>/upload-<BEST>
mkdir -p $UPLOAD_DIR
cp $RUN_DIR/exports/global_step_<BEST>/policy/* $UPLOAD_DIR/
ls $UPLOAD_DIR/      # *.safetensors, config.json, tokenizer files all at root

4. Derive the size suffix from exported weights

The repo is laion/<run_name>-<BEST>-<size>B. Compute <size>B from config.json and safetensor parameters; never parse it from the run or base-model name:

/leonardo_work/AIFAC_5C0_290/bfeuer00/miniforge3/envs/otagent/bin/python - "$UPLOAD_DIR" <<'PY'
import json, sys, glob, os
d = sys.argv[1]
cfg = json.load(open(os.path.join(d, "config.json")))
H = cfg.get("hidden_size"); L = cfg.get("num_hidden_layers"); V = cfg.get("vocab_size")
# 1) Exact: sum param counts from the safetensors index (preferred).
idx = glob.glob(os.path.join(d, "*.safetensors.index.json"))
total = None
if idx:
    wm = json.load(open(idx[0])).get("metadata", {})
    total = wm.get("total_parameters") or wm.get("total_size")  # total_size = BYTES, not params
    if total and "total_parameters" not in wm:
        total = None  # only had byte size; fall through to header sum
if total is None:
    # Sum tensor numels from each shard header (no torch load needed).
    import struct
    total = 0
    for f in glob.glob(os.path.join(d, "*.safetensors")):
        with open(f, "rb") as fh:
            n = struct.unpack("<Q", fh.read(8))[0]
            hdr = json.loads(fh.read(n))
        for name, meta in hdr.items():
            if name == "__metadata__": continue
            shp = meta.get("shape", [])
            cnt = 1
            for s in shp: cnt *= s
            total += cnt
# 2) Cross-check with the transformer estimate 12*L*H^2 + embeddings(2*V*H).
est = 12 * L * H * H + 2 * (V or 0) * H if (H and L) else None
print(f"config: hidden_size={H} layers={L} vocab={V}")
(f  total  )
 est: (f)
b = total or est
(f)
PY

Use summed parameters; the estimate is a cross-check. If they differ by >~15%, prefer summed and note it. Round to the nearest whole G (9.7B → 10B).

5. Copy the launch config + scan for secrets

cp $WORK/code/OpenThoughts-Agent/hpc/skyrl_standard/leonardo/run_delphi_math_rl.sh   $UPLOAD_DIR/rl_run_script.sh 2>/dev/null
cp $WORK/code/OpenThoughts-Agent/hpc/skyrl_standard/leonardo/sbatch_delphi_math_rl.sh $UPLOAD_DIR/rl_sbatch.sh    2>/dev/null
trufflehog filesystem $UPLOAD_DIR --no-update    # if installed; else the grep fallback:
grep -rIE '(sk-[a-zA-Z0-9]{20,}|AKIA[0-9A-Z]{16}|ghp_[a-zA-Z0-9]{36}|hf_[a-zA-Z0-9]{34}|eyJ[a-zA-Z0-9._-]+)' $UPLOAD_DIR

Remove/redact anything found before upload (HF runs TruffleHog post-upload; catch it first).

6. Stage the metrics alongside the model, then upload — laion/<run_name>-<BEST>-<size>B

Fold metric outputs into the upload dir (runs are WANDB_MODE=offline, so these are the only persistent record):

mkdir -p $UPLOAD_DIR/training_logs
cp $OUT/metrics.csv $OUT/vllm_metrics.csv $OUT/report.md $OUT/reward_plot.png $UPLOAD_DIR/training_logs/ 2>/dev/null
cp $WORK/code/OpenThoughts-Agent/hpc/skyrl_standard/leonardo/<JobName>_*.out      $UPLOAD_DIR/training_logs/    2>/dev/null

Upload via the Leonardo sbatch-tunnel (NOT a login-node hf upload — dies at ~100 s). Use the sbatch template in .agents/ops/leonardo/ops.md ("sbatch template for HF upload"); point cd at $UPLOAD_DIR and the command at:

$CMD_PREFIX /leonardo_work/AIFAC_5C0_290/bfeuer00/miniforge3/envs/otagent/bin/hf upload \
    laion/<run_name>-<BEST>-<size>B . . --repo-type=model

7. Register in the DB (--training-type RL) — ONLY if the series is DB-registerable

Confirm the cell's series is DB-registerable BEFORE registering. Several standard-RL series are HF-upload-only (the Delphi scaling-laws RL cells, like the SFT grid). If it's ambiguous, STOP and flag — do NOT auto-register a no-DB series (it pollutes the registry; artifacts are consumed by an eval grid). Honor enable_db_registration: false and documented no-DB series (crud-otagent-supabase → "Per-series exceptions").

Base model (--base-model) = the checkpoint the RL trained FROM — the sbatch/run-script MODEL_PATH (read from the .out launch line trainer.policy.model.path=… or the sbatch MODEL_PATH= token), e.g. laion/delphi-1e21-…-sft. --base-model resolves to the base_model_id self-FK — it must name an already-registered models row (register/point it first if not, per crud-otagent-supabase).

Standard GRPO does not auto-push a canonical laion/<run> duplicate, so there's usually no auto-row to delete. Still run the cross-user FK pre-check before any delete/mutate; restrict writes to rows you own:

other_users_fk = (c.table("sandbox_jobs").select("id,username,model_id")
    .eq("model_id", stray_row_id).neq("username", os.environ.get("USER","bfeuer00")).execute())
if other_users_fk.data:
    print(f"SKIPPING delete — {len(other_users_fk.data)} other-user rows FK'd; surface + leave it.")

Then register (--training-type RL REQUIRED — defaults to SFT):

python scripts/database/manual_db_push.py \
  --hf-model-id laion/<run_name>-<BEST>-<size>B \
  --base-model laion/delphi-…-sft \
  --dataset-name <rlvr_math|gsm8k|aime|…> \
  --training-type RL

8. Clean up the run dir

After upload and applicable DB registration are confirmed on HF, free disk. Use detached rm -rf; never du or find first. Verify inode reclaim afterward (jutil project dataquota -p <project> / df -i):

nohup rm -rf $WORK/rl_ckpts/<RUN_NAME> >/dev/null 2>&1 &   # detached; keep $UPLOAD_DIR until HF verified

Keep $WORK/rl_cleanup/<RUN_NAME>/ (staging + metrics) until you've confirmed the HF repo lists the weights + training_logs/, then remove it too.

9. Fire the Delphi eval suite on the post-RL checkpoint

Score the uploaded ckpt on the lab's fixed downstream suite. DEFERS to eval-standard-launch → "§5b. Evaluate a (post-RL) checkpoint on the Delphi eval suite" — do not re-document it here. That section runs the canonical delphi_eval.sbatch (MATH500 / AIME24 10-seed / gsm8k, pass@1, temp 0.7, delphi_v0 template, STAGE=rl) on Leonardo against the uploaded laion/ repo. One-line invocation (pre-cache the repo on the login node first, per that skill §3):

RUN=<run_name>-<BEST>-<size>B   # the repo §6 just published
sbatch --job-name="delphi-eval-$RUN" \
  /leonardo_work/AIFAC_5C0_290/bfeuer00/experiments/delphi-eval/delphi_eval.sbatch laion/$RUN $RUN rl

Add the submitted row to main_rl_evals/SCORES.md (🚀 eval submitted + job id); harvest via eval-standard-cleanup. HF-upload-only, NEVER DB — same series rule as §7.