#!/bin/bash
#SBATCH --job-name=smalltest
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuH200x8
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --gres=gpu:4
#SBATCH --mem=200G
#SBATCH --time=02:00:00
#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/smalltest-%j.out
#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/smalltest-%j.err

# Small-scale test: 5 profiles, 5 sessions, all 6 methods
# Full settings (70B user sim, 8B agent) but fewer questions

cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents
source /u/yurenh2/miniforge3/etc/profile.d/conda.sh
conda activate eval
export HF_HOME=/projects/bfqt/users/yurenh2/hf_cache/huggingface
export PYTHONPATH="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/src:$PYTHONPATH"

PROFILE_PATH="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents/data/complex_profiles_v2/profiles_200.jsonl"
AGENT_MODEL="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
USER_MODEL="meta-llama/Llama-3.1-70B-Instruct"

echo "=== Small-scale Test: All 6 Methods ==="
echo "Settings: 5 profiles, 5 sessions each, max 15 turns"
echo "User simulator: $USER_MODEL (70B)"
echo "Agent: $AGENT_MODEL (8B)"
date

# Start vLLM servers
# User simulator on GPUs 0,1 (70B, TP=2)
CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
    --model $USER_MODEL \
    --port 8004 --tensor-parallel-size 2 --gpu-memory-utilization 0.90 \
    --max-model-len 16384 --dtype bfloat16 --download-dir $HF_HOME &

# Agent on GPUs 2,3 (8B, TP=2, lower memory for embedding/reranker)
CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
    --model $AGENT_MODEL \
    --port 8003 --tensor-parallel-size 2 --gpu-memory-utilization 0.45 \
    --max-model-len 16384 --dtype bfloat16 &

# Wait for servers
echo "Waiting for vLLM servers..."
for i in {1..200}; do
    if curl -s http://localhost:8004/health > /dev/null 2>&1; then
        echo "User simulator (8004) ready after $((i*5)) seconds"
        break
    fi
    sleep 5
done
for i in {1..60}; do
    if curl -s http://localhost:8003/health > /dev/null 2>&1; then
        echo "Agent (8003) ready after $((i*5)) seconds"
        break
    fi
    sleep 5
done
echo "Both vLLM servers ready"
sleep 10

# Run all 6 methods sequentially with small scale
for METHOD in vanilla contextual reflection all_memory rag rag_vector; do
    echo ""
    echo "=== Testing method: $METHOD ==="
    date

    python scripts/run_experiments.py --methods $METHOD \
        --datasets math-hard --n-profiles 5 --n-sessions 5 --max-turns 15 \
        --use-vllm --no-batch-processing --parallel-profiles 5 \
        --output-dir ../results/smalltest --profile-path $PROFILE_PATH

    if [ $? -eq 0 ]; then
        echo "Method $METHOD: SUCCESS"
    else
        echo "Method $METHOD: FAILED"
    fi
done

echo ""
echo "=== Small-scale test complete ==="
date

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