code/scripts/queue_grpo.sh

#!/usr/bin/env bash
# Chain: GRPO train → z-ablation eval on final → push to HF under grpo/
#
# Run as:
#   BLT_HF_TOKEN=<hf_token> nohup bash queue_grpo.sh > /home/ubuntu/work/queue_grpo.log 2>&1 &
set -uo pipefail

REPO="LauraGG/blt-reasoner-pilot1"
OUT="/home/ubuntu/work/blt_grpo1"
CFG="/home/ubuntu/experiments/blt_reasoner/configs/grpo_from_step12000.json"
LOG="/home/ubuntu/work/queue_grpo.log"

log() { echo "[$(date +%T)] $*" | tee -a "$LOG"; }

mkdir -p "$OUT"
cd /home/ubuntu
export TOKENIZERS_PARALLELISM=false TRANSFORMERS_NO_ADVISORY_WARNINGS=1 HF_HUB_DISABLE_PROGRESS_BARS=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

log "=========================================="
log "GRPO PHASE C: train from pilot ckpt-step12000"
log "=========================================="
python3 -u -m experiments.blt_reasoner.grpo_train --config "$CFG" \
    > "$OUT/train.log" 2>&1
rc=$?
log "GRPO train exit=$rc"
if [ $rc -ne 0 ]; then
    log "GRPO train FAILED; pushing partial state and exiting"
fi

# Always try to push whatever we have
log "running n=100 K=16 z-ablation on GRPO final ckpt"
python3 -u -m experiments.blt_reasoner.eval \
    --ckpt "$OUT/final" --config "$CFG" \
    --n 100 --K 16 --max_new_tokens 192 --temperature 0.0 \
    --out "$OUT/final/ablation_K16_n100.json" \
    > "$OUT/eval.log" 2>&1
log "eval exit=$?"

log "pushing to HF under grpo/"
python3 - <<PY
import os
from huggingface_hub import HfApi
token = os.environ.get("BLT_HF_TOKEN", "").strip()
assert token.startswith("hf_"), "BLT_HF_TOKEN missing"
api = HfApi(token=token)
api.upload_folder(
    folder_path="$OUT",
    path_in_repo="grpo",
    repo_id="$REPO", repo_type="model",
    commit_message="GRPO Phase C: trained policy + final ablation",
    ignore_patterns=["*.tmp"],
)
print("[push] done")
PY
log "queue_grpo.sh DONE"