#!/bin/bash
#SBATCH --job-name=rag_v2
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuH200x8
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --gres=gpu:h200:4
#SBATCH --mem=200G
#SBATCH --time=2:00:00
#SBATCH --output=rag_v2_%j.out
#SBATCH --error=rag_v2_%j.err

set -e
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

MODEL_70B="meta-llama/Llama-3.1-70B-Instruct"
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=0,1 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_70B --port 8004 --tensor-parallel-size 2 \
    --gpu-memory-utilization 0.95 --max-model-len 16384 \
    --download-dir $HF_HOME --dtype bfloat16 --disable-log-requests &

CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_8B --port 8003 --tensor-parallel-size 2 \
    --gpu-memory-utilization 0.40 --max-model-len 16384 \
    --dtype bfloat16 --disable-log-requests &

for i in $(seq 1 120); do
    R1=$(curl -s http://localhost:8004/health > /dev/null 2>&1 && echo 1 || echo 0)
    R2=$(curl -s http://localhost:8003/health > /dev/null 2>&1 && echo 1 || echo 0)
    [ "$R1" = "1" ] && [ "$R2" = "1" ] && break
    sleep 3
done

cd collaborativeagents/scripts
python run_experiments.py \
    --methods rag \
    --datasets math-hard \
    --n-profiles 5 \
    --n-sessions 20 \
    --use-vllm \
    --vllm-user-url http://localhost:8004/v1 \
    --vllm-agent-url http://localhost:8003/v1 \
    --parallel-profiles 5 \
    --profile-path ../data/complex_profiles_v2/profiles_200.jsonl \
    --output-dir ../results/rag_v2_$(date +%Y%m%d_%H%M%S)

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