#!/bin/bash
#SBATCH --job-name=batch_exp
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuH200x8
#SBATCH --gres=gpu:4
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --mem=256G
#SBATCH --time=24:00:00
#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/batch_exp-%j.out
#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/batch_exp-%j.err

# Full experiment: Batch-processable methods (vanilla, all_memory)
# 200 profiles × 30 sessions = 6,000 sessions per method
# Using turn-synchronous batch processing (paper's approach)

set -e

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="${PWD}:${PWD}/../src:${PYTHONPATH}"

MODEL_8B="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
PORT_USER=8004
PORT_AGENT=8003
PROFILE_PATH="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents/data/complex_profiles_v2/profiles_200.jsonl"

echo "============================================"
echo "Full Experiment: Batch Methods"
echo "============================================"
echo "Methods: vanilla, all_memory"
echo "Profiles: 200"
echo "Sessions/profile: 30"
echo "Total: 6,000 sessions per method"
echo ""
date
nvidia-smi --query-gpu=index,name,memory.total --format=csv
echo ""

# Kill any existing servers
pkill -f "vllm.entrypoints" 2>/dev/null || true
sleep 2

# Start vLLM servers with optimized settings
echo "Starting 8B user simulator (GPU 0-1, TP=2)..."
CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_8B \
    --port $PORT_USER \
    --tensor-parallel-size 2 \
    --gpu-memory-utilization 0.90 \
    --max-model-len 8192 \
    --disable-log-requests \
    --dtype bfloat16 \
    --max-num-seqs 256 &
SERVER_USER_PID=$!

echo "Starting 8B agent (GPU 2-3, TP=2)..."
CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_8B \
    --port $PORT_AGENT \
    --tensor-parallel-size 2 \
    --gpu-memory-utilization 0.90 \
    --max-model-len 8192 \
    --disable-log-requests \
    --dtype bfloat16 \
    --max-num-seqs 256 &
SERVER_AGENT_PID=$!

echo "Waiting for servers (may take 5-10 min for CUDA graph compilation)..."
for i in $(seq 1 200); do
    READY_USER=$(curl -s http://localhost:$PORT_USER/health > /dev/null 2>&1 && echo 1 || echo 0)
    READY_AGENT=$(curl -s http://localhost:$PORT_AGENT/health > /dev/null 2>&1 && echo 1 || echo 0)
    if [ "$READY_USER" = "1" ] && [ "$READY_AGENT" = "1" ]; then
        echo "Both servers ready after $((i*3))s"
        break
    fi
    if [ $((i % 20)) -eq 0 ]; then
        echo "  Still waiting... ($((i*3))s)"
    fi
    sleep 3
done

if ! curl -s http://localhost:$PORT_USER/health > /dev/null; then
    echo "ERROR: User server not healthy"; kill $SERVER_USER_PID $SERVER_AGENT_PID 2>/dev/null; exit 1
fi
if ! curl -s http://localhost:$PORT_AGENT/health > /dev/null; then
    echo "ERROR: Agent server not healthy"; kill $SERVER_USER_PID $SERVER_AGENT_PID 2>/dev/null; exit 1
fi
echo "Both servers healthy"
echo ""

# Run batch experiment (only vanilla is truly stateless)
for METHOD in vanilla; do
    echo "============================================"
    echo "Running: $METHOD (BATCH processing)"
    echo "============================================"
    START=$(date +%s)

    python scripts/run_experiments.py \
        --methods $METHOD \
        --datasets math-hard,math-500,bigcodebench \
        --n-profiles 200 \
        --n-sessions 30 \
        --max-turns 15 \
        --use-vllm \
        --batch-size 50 \
        --parallel-profiles 50 \
        --output-dir ../results/full_experiment_h200 \
        --profile-path "$PROFILE_PATH"

    END=$(date +%s)
    ELAPSED=$((END-START))
    echo ""
    echo "$METHOD completed in ${ELAPSED}s"
    THROUGHPUT=$((6000 * 3600 / ELAPSED))
    echo "Throughput: ${THROUGHPUT} sessions/hr"
    echo ""
done

# Cleanup
echo "Cleaning up..."
kill $SERVER_USER_PID $SERVER_AGENT_PID 2>/dev/null || true

echo ""
echo "============================================"
echo "BATCH EXPERIMENT COMPLETE"
echo "============================================"
date