#!/bin/bash
#SBATCH --job-name=vl
#SBATCH --partition=gpu-large
#SBATCH --qos=batch-long
#SBATCH --gres=gpu:h200:2
#SBATCH --cpus-per-task=8
#SBATCH --mem=180G
#SBATCH --time=03:00:00
#SBATCH --output=/weka/s225250685/mats-tist/slurm_logs/gen_bird_yes_only_variations_%j.out

# Generate K=7 wrong-SQL variations for the 3128 BIRD-train YES-only questions
# using Qwen-2.5-72B (TP=2).

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

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

LOG=/weka/s225250685/mats-tist/slurm_logs/gen_bird_yes_only_variations_${SLURM_JOB_ID}.log
kill_vllm() { pkill -9 -f "vllm serve" 2>/dev/null || true; sleep 5; }
trap kill_vllm EXIT
wait_url() { for i in {1..720}; do curl --noproxy '*' -fs "$1" >/dev/null 2>&1 && return 0; sleep 5; done; return 1; }

$VLLM serve "Qwen/Qwen2.5-72B-Instruct" \
    --served-model-name teacher --port 8200 \
    --dtype bfloat16 --tensor-parallel-size 2 \
    --gpu-memory-utilization 0.92 --max-model-len 4096 \
    --max-num-seqs 32 --enforce-eager --disable-log-requests \
    > "${LOG}.serve" 2>&1 &
wait_url http://localhost:8200/v1/models || { echo "teacher failed" | tee -a "$LOG"; exit 1; }

$PY scripts/gen_synsql_wrong_variations.py \
    --input data/external/bird_yes_only.jsonl \
    --out data/external/bird_yes_only_with_variations.jsonl \
    --teacher_host http://localhost:8200 \
    --k_wrong 7 --threads 32 --resume \
    2>&1 | tee -a "${LOG}.gen"

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