#!/bin/bash
#SBATCH --job-name=rag
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuA100x4
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=16
#SBATCH --gres=gpu:nvidia_a100:2
#SBATCH --mem=128G
#SBATCH --time=2:00:00
#SBATCH --output=rag_%j.out
#SBATCH --error=rag_%j.err

cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model
source /u/yurenh2/miniforge3/etc/profile.d/conda.sh
conda activate eval

export HF_HOME=/projects/bfqt/users/yurenh2/hf_cache/huggingface
export PYTHONPATH="${PWD}/src:${PWD}/collaborativeagents:${PYTHONPATH}"
export NCCL_P2P_DISABLE=1

set -a; source .env; set +a
pip install --quiet openai python-dotenv json-repair

MODEL_8B="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"

pkill -f "vllm.entrypoints" 2>/dev/null || true
sleep 2

CUDA_VISIBLE_DEVICES=0 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_8B --port 8003 --tensor-parallel-size 1 \
    --gpu-memory-utilization 0.5 --max-model-len 8192 \
    --dtype bfloat16 --disable-log-requests &

for i in $(seq 1 90); do
    curl -s http://localhost:8003/health > /dev/null 2>&1 && break
    sleep 2
done
echo "vLLM ready."

cd collaborativeagents/scripts
python run_experiments.py \
    --methods rag \
    --datasets math-hard \
    --n-profiles 20 \
    --n-sessions 5 \
    --max-turns 8 \
    --use-vllm \
    --use-openai-user \
    --openai-user-model gpt-5 \
    --reward-mode llm \
    --vllm-agent-url http://localhost:8003/v1 \
    --parallel-profiles 20 \
    --profile-path ../data/complex_profiles_v2/profiles_200.jsonl \
    --output-dir ../results/gpt_user_scale_rag_$(date +%Y%m%d_%H%M%S)

pkill -f "vllm.entrypoints" 2>/dev/null || true