#!/bin/bash
#SBATCH --job-name=vl
#SBATCH --partition=gpu-large
#SBATCH --qos=batch-long
#SBATCH --gres=gpu:2
#SBATCH --cpus-per-task=8
#SBATCH --mem=200G
#SBATCH --time=08:00:00
#SBATCH --output=/weka/s225250685/mats-tist/slurm_logs/gen_validator_qwen72b_%j.out

# Serve Qwen-2.5-72B-Instruct (bf16, TP=2 across 2 H100s) as teacher.
# Reads:  data/planner_3B_greedy_bird_train.jsonl  (depends on 88586)
# Writes: data/hf_val_sel_paper_v1, data/hf_val_cond_paper_v1

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
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

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; }

echo "[$(date)] Serving Qwen2.5-72B-Instruct (bf16, TP=2) on port 8200..."
$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 8192 \
    --enforce-eager \
    > /tmp/vllm_teacher_$$ 2>&1 &
wait_url http://localhost:8200/v1/models
echo "[$(date)] Teacher ready"

echo "[$(date)] Generating validator SFT data..."
$PY scripts/gen_validator_sft_qwen72b.py \
    --input data/planner_3B_greedy_bird_train.jsonl \
    --out_sel data/hf_val_sel_paper_v1 \
    --out_cond data/hf_val_cond_paper_v1 \
    --teacher_host http://localhost:8200 \
    --temperature 0.3 --batch_size 16

echo "[$(date)] DONE"