#!/bin/bash
#SBATCH --job-name=vl
#SBATCH --partition=gpu-large
#SBATCH --qos=batch-long
#SBATCH --gres=gpu:h200:1
#SBATCH --cpus-per-task=8
#SBATCH --mem=120G
#SBATCH --time=08:00:00
#SBATCH --output=/weka/s225250685/mats-tist/slurm_logs/eval_selector_v5_%j.out

# Phase 3 — Pairwise tournament eval for selector v5.
# Serve the trained Llama-3B selector with vLLM, then run pairwise tournament
# eval on paper-format K=8 dev rollouts.

set -u
cd /weka/s225250685/mats-tist
set -a; source /weka/s225250685/mats-tist/.env; set +a
export HF_HOME=/weka/s225250685/Huggingface HF_HUB_CACHE=/weka/s225250685/Huggingface/hub
export PYTHONNOUSERSITE=1 NO_PROXY=localhost,127.0.0.1 no_proxy=localhost,127.0.0.1
export PYTHONPATH=/weka/s225250685/mats-tist TOKENIZERS_PARALLELISM=false
export DB_EXEC_API_DISABLE=1

PY=/weka/s225250685/conda-envs/handbook/bin/python
VLLM=/weka/s225250685/conda-envs/handbook/bin/vllm

SELECTOR_CKPT=${SELECTOR_CKPT:-alignment-handbook/output/selector-llama3b-v5-pairwise-rich}
LABEL=${LABEL:-v5_paper_SFT_VF}
DEV_FILE=${DEV_FILE:-eval_results/paper_SFT_VF_passAt8_bird_dev.jsonl}

LOG=/weka/s225250685/mats-tist/slurm_logs/eval_selector_v5_${SLURM_JOB_ID}.log
: > "$LOG"

kill_vllm() { pkill -9 -f "vllm serve" 2>/dev/null || true; sleep 5; }
trap kill_vllm EXIT
wait_url() {
    for i in {1..360}; do
        curl --noproxy '*' -fs "$1" >/dev/null 2>&1 && return 0
        sleep 5
    done
    return 1
}

echo "[$(date)] Serving selector: $SELECTOR_CKPT" | tee -a "$LOG"
$VLLM serve "$SELECTOR_CKPT" \
    --served-model-name selector --port 8103 \
    --dtype bfloat16 \
    --gpu-memory-utilization 0.85 \
    --max-model-len 4096 \
    --enforce-eager \
    --disable-log-requests \
    > "$LOG.serve" 2>&1 &

if ! wait_url http://localhost:8103/v1/models; then
    echo "[$(date)] selector failed to come up" | tee -a "$LOG"
    tail -200 "$LOG.serve" | tee -a "$LOG"
    exit 1
fi
echo "[$(date)] selector ready" | tee -a "$LOG"

echo "[$(date)] Running pairwise tournament on $DEV_FILE..." | tee -a "$LOG"
$PY scripts/compute_bestofn_pairwise_v5.py \
    "$DEV_FILE" "$LABEL" \
    --selector_host http://localhost:8103 \
    --model_name selector \
    --threads 16 \
    --out_dir eval_results \
    2>&1 | tee -a "$LOG.eval"

echo "[$(date)] DONE" | tee -a "$LOG"
kill_vllm