#!/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:4
#SBATCH --mem=250G
#SBATCH --time=03:00:00
#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
# 5 profiles, 15 sessions each (more sessions to test accumulation)
# Compare: vanilla, rag, rag_vector

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

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 ==="
echo "Key change: Starting with EMPTY memory store"
echo "  - RAG will accumulate memories during evaluation"
echo "  - Each user builds their own memory basket from scratch"
echo ""
echo "Settings: 5 profiles, 15 sessions each"
echo "User simulator: $USER_MODEL (70B)"
echo "Agent: $AGENT_MODEL (8B)"
date
nvidia-smi --query-gpu=index,name,memory.total --format=csv

# Clear empty store before each run to ensure fresh start
echo ""
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
echo "Memory store cleared."

# Start vLLM servers with adjusted memory allocation
echo ""
echo "Starting 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 0.85 \
    --max-model-len 16384 --dtype bfloat16 --download-dir $HF_HOME \
    --disable-log-requests &
USER_PID=$!

# Agent on GPUs 2,3 (8B, TP=2) - reduced memory for embedding/reranker headroom
CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
    --model $AGENT_MODEL \
    --port 8003 --tensor-parallel-size 2 --gpu-memory-utilization 0.40 \
    --max-model-len 16384 --dtype bfloat16 \
    --disable-log-requests &
AGENT_PID=$!

# Wait for servers
echo "Waiting for vLLM servers (may take 5-10 min)..."
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

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)"

# Run methods sequentially (each starts with fresh empty memory)
for METHOD in vanilla rag rag_vector; do
    echo ""
    echo "============================================"
    echo "Testing method: $METHOD"
    echo "============================================"

    # Clear memory store before each method for fair comparison
    > /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
    echo "Memory store cleared for $METHOD"

    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))

    # Show memory accumulation stats for RAG methods
    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

    if [ $? -eq 0 ]; then
        echo "Method $METHOD: SUCCESS (${ELAPSED}s)"
    else
        echo "Method $METHOD: FAILED after ${ELAPSED}s"
    fi
done

echo ""
echo "============================================"
echo "RAG Empty Memory Test Complete"
echo "============================================"
echo "Results saved to: $OUTPUT_DIR"
date

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