#!/bin/bash
#SBATCH --job-name=ctrl_test
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuH200x8-interactive
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --gres=gpu:4
#SBATCH --mem=200G
#SBATCH --time=00:45:00
#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/ctrl_test-%j.out
#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/ctrl_test-%j.err

# Controlled Test: Same user profile, same questions, 3 methods
# Tests:
# 1. Stronger user enforcement prompts
# 2. Memory retrieval debug output
# 3. Comparison across 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"

# Use first profile only for controlled comparison
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"
MEMORY_STORE="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store"

echo "=== Controlled Comparison Test ==="
echo "Same user profile (1st), same 15 questions, 3 methods"
echo "Testing: stronger enforcement + retrieval debug"
date
nvidia-smi --query-gpu=index,name,memory.total --format=csv

# Start vLLM servers
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 &

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 &

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 ready after $((i*5))s"
        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 ready after $((i*5))s"
        break
    fi
    sleep 5
done
sleep 5

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

# Run each method with SAME user (1 profile, 15 sessions)
for METHOD in vanilla rag rag_vector; do
    echo ""
    echo "============================================"
    echo "Testing: $METHOD"
    echo "============================================"

    # Clear memory store before each method (fresh start)
    > ${MEMORY_STORE}/memory_cards.jsonl
    rm -f ${MEMORY_STORE}/memory_embeddings.npy
    echo "Memory store cleared"

    date
    python scripts/run_experiments.py --methods $METHOD \
        --datasets math-hard --n-profiles 1 --n-sessions 15 --max-turns 15 \
        --use-vllm --no-batch-processing --parallel-profiles 1 \
        --output-dir $OUTPUT_DIR --profile-path $PROFILE_PATH

    echo "Method $METHOD completed"

    # Show memory count for rag methods
    if [ "$METHOD" != "vanilla" ]; then
        echo "Final memory cards: $(wc -l < ${MEMORY_STORE}/memory_cards.jsonl)"
    fi
done

echo ""
echo "=== Done ==="
date

# Generate comparison summary
python3 << 'EOF'
import json
import os
from pathlib import Path

output_base = sorted(Path("../results").glob("controlled_test_*"))[-1]
print(f"\n=== Comparison Summary ===\n")
print(f"Results dir: {output_base}")

methods = ["vanilla", "rag", "rag_vector"]
results = {}

for subdir in output_base.iterdir():
    if subdir.is_dir():
        for method in methods:
            result_file = subdir / method / "results.json"
            if result_file.exists():
                with open(result_file) as f:
                    results[method] = json.load(f)
                break

if results:
    print(f"\n{'Metric':<25} {'vanilla':<12} {'rag':<12} {'rag_vector':<12}")
    print("-" * 60)

    for method in methods:
        if method not in results:
            continue
        data = results[method]
        task_succ = sum(r['metrics']['task_success'] for r in data) / len(data)
        avg_turns = sum(r['metrics']['total_turns'] for r in data) / len(data)
        avg_enf = sum(r['metrics']['enforcement_count'] for r in data) / len(data)

        if method == methods[0]:
            print(f"{'Task Success':<25} {task_succ:<12.1%} ", end="")
        else:
            print(f"{task_succ:<12.1%} ", end="")
    print()

    for method in methods:
        if method not in results:
            continue
        data = results[method]
        avg_turns = sum(r['metrics']['total_turns'] for r in data) / len(data)
        if method == methods[0]:
            print(f"{'Avg Turns':<25} {avg_turns:<12.1f} ", end="")
        else:
            print(f"{avg_turns:<12.1f} ", end="")
    print()

    for method in methods:
        if method not in results:
            continue
        data = results[method]
        avg_enf = sum(r['metrics']['enforcement_count'] for r in data) / len(data)
        if method == methods[0]:
            print(f"{'Avg Enforcement':<25} {avg_enf:<12.1f} ", end="")
        else:
            print(f"{avg_enf:<12.1f} ", end="")
    print()

    # Session-by-session comparison
    print(f"\n=== Session-by-Session Turns ===")
    print(f"{'Session':<10} {'vanilla':<12} {'rag':<12} {'rag_vector':<12}")
    print("-" * 50)
    for i in range(min(15, len(results.get('vanilla', [])))):
        print(f"{i+1:<10} ", end="")
        for method in methods:
            if method in results and i < len(results[method]):
                turns = results[method][i]['metrics']['total_turns']
                print(f"{turns:<12} ", end="")
        print()
EOF

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