#!/bin/bash
#SBATCH --job-name=rag_empty
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuH200x8
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --gres=gpu:8
#SBATCH --mem=400G
#SBATCH --time=04:00:00
#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/rag_empty-%j.out
#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/rag_empty-%j.err

# Test RAG with EMPTY memory store - start fresh and accumulate
# Using 8 GPUs with TP=4 for 70B to avoid CUDA errors

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"
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export NCCL_P2P_DISABLE=1

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 "=== RAG Empty Memory Store Test v2 ==="
echo "Using 8 GPUs: TP=4 for 70B user sim, TP=2 for 8B agent"
date
nvidia-smi --query-gpu=index,name,memory.total --format=csv

# Clear empty store
echo "Clearing empty memory store..."
> /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store/memory_cards.jsonl
rm -f /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store/memory_embeddings.npy

# Start vLLM servers
echo "Starting vLLM servers..."

# User simulator on GPUs 0-3 (70B, TP=4)
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m vllm.entrypoints.openai.api_server \
    --model $USER_MODEL \
    --port 8004 --tensor-parallel-size 4 --gpu-memory-utilization 0.90 \
    --max-model-len 16384 --dtype bfloat16 --download-dir $HF_HOME \
    --disable-log-requests &
USER_PID=$!

# Agent on GPUs 4-5 (8B, TP=2)
CUDA_VISIBLE_DEVICES=4,5 python -m vllm.entrypoints.openai.api_server \
    --model $AGENT_MODEL \
    --port 8003 --tensor-parallel-size 2 --gpu-memory-utilization 0.40 \
    --max-model-len 16384 --dtype bfloat16 \
    --disable-log-requests &
AGENT_PID=$!

# Wait for servers
echo "Waiting for vLLM servers..."
for i in {1..300}; 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..120}; 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

if ! curl -s http://localhost:8004/health > /dev/null 2>&1; then
    echo "ERROR: User server not healthy"
    kill $USER_PID $AGENT_PID 2>/dev/null
    exit 1
fi
if ! curl -s http://localhost:8003/health > /dev/null 2>&1; then
    echo "ERROR: Agent server not healthy"
    kill $USER_PID $AGENT_PID 2>/dev/null
    exit 1
fi
echo "Both vLLM servers ready"
sleep 5

OUTPUT_DIR="../results/rag_empty_test_$(date +%Y%m%d_%H%M%S)"

for METHOD in vanilla rag rag_vector; do
    echo ""
    echo "============================================"
    echo "Testing method: $METHOD"
    echo "============================================"

    # Clear memory store before each method
    > /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store/memory_cards.jsonl
    rm -f /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store/memory_embeddings.npy

    date
    START=$(date +%s)

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

    END=$(date +%s)
    ELAPSED=$((END-START))

    if [[ "$METHOD" == "rag" || "$METHOD" == "rag_vector" ]]; then
        CARD_COUNT=$(wc -l < /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store/memory_cards.jsonl 2>/dev/null || echo 0)
        echo "Memory cards accumulated: $CARD_COUNT"
    fi

    echo "Method $METHOD: completed in ${ELAPSED}s"
done

echo ""
echo "=== Test Complete ==="
echo "Results: $OUTPUT_DIR"
date

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