#!/bin/bash
#SBATCH --job-name=vllm_bench
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuA100x4
#SBATCH --nodes=1
#SBATCH --gpus-per-node=2
#SBATCH --time=02:00:00
#SBATCH --mem=128G
#SBATCH --output=slurm/logs/vllm_bench_70b_8b_%j.out
#SBATCH --error=slurm/logs/vllm_bench_70b_8b_%j.err

# Realistic benchmark: 70B AWQ user simulator + 8B agent
# Tests actual conversation throughput with both models
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

echo "=== Job Info ==="
echo "Job ID: $SLURM_JOB_ID"
echo "Node: $SLURM_NODELIST"
date

echo ""
echo "=== GPU Info ==="
nvidia-smi --query-gpu=index,name,memory.total,memory.free --format=csv

# Download AWQ 70B model if not complete
echo ""
echo "=== Ensuring AWQ 70B Model is Downloaded ==="
python -c "
from huggingface_hub import snapshot_download
import os
os.environ['HF_HOME'] = '/projects/bfqt/users/yurenh2/hf_cache/huggingface'
print('Checking/downloading hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4...')
path = snapshot_download('hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4')
print(f'Model ready at: {path}')
"

MODEL_70B_AWQ="hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4"
MODEL_8B="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
PORT_70B=8004
PORT_8B=8003

echo ""
echo "============================================"
echo "Starting 70B AWQ vLLM Server (GPU 0)"
echo "============================================"
CUDA_VISIBLE_DEVICES=0 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_70B_AWQ \
    --port $PORT_70B \
    --gpu-memory-utilization 0.90 \
    --max-model-len 4096 \
    --disable-log-requests \
    --quantization awq \
    --dtype float16 &
SERVER_70B_PID=$!
echo "70B Server PID: $SERVER_70B_PID"

echo ""
echo "============================================"
echo "Starting 8B vLLM Server (GPU 1)"
echo "============================================"
CUDA_VISIBLE_DEVICES=1 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_8B \
    --port $PORT_8B \
    --gpu-memory-utilization 0.90 \
    --max-model-len 4096 \
    --disable-log-requests \
    --dtype bfloat16 &
SERVER_8B_PID=$!
echo "8B Server PID: $SERVER_8B_PID"

echo ""
echo "Waiting for servers to start..."

# Wait for 70B (may take 3-5 minutes)
for i in $(seq 1 120); do
    if curl -s http://localhost:$PORT_70B/health > /dev/null 2>&1; then
        echo "70B Server ready after $((i*3)) seconds"
        break
    fi
    if [ $((i % 20)) -eq 0 ]; then
        echo "  Waiting for 70B... ($((i*3)) seconds)"
    fi
    sleep 3
done

# Wait for 8B
for i in $(seq 1 60); do
    if curl -s http://localhost:$PORT_8B/health > /dev/null 2>&1; then
        echo "8B Server ready after $((i*2)) seconds"
        break
    fi
    sleep 2
done

# Check both servers
echo ""
if ! curl -s http://localhost:$PORT_70B/health > /dev/null 2>&1; then
    echo "ERROR: 70B server failed to start"
    kill $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null
    exit 1
fi
echo "✓ 70B server healthy"

if ! curl -s http://localhost:$PORT_8B/health > /dev/null 2>&1; then
    echo "ERROR: 8B server failed to start"
    kill $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null
    exit 1
fi
echo "✓ 8B server healthy"

echo ""
echo "=== vLLM Server Info ==="
echo "70B model:"
curl -s http://localhost:$PORT_70B/v1/models | python -m json.tool 2>/dev/null | head -10
echo ""
echo "8B model:"
curl -s http://localhost:$PORT_8B/v1/models | python -m json.tool 2>/dev/null | head -10

echo ""
echo "============================================"
echo "Test 1: Individual Model Throughput"
echo "============================================"

echo ""
echo "--- 70B AWQ Sequential (10 requests) ---"
python scripts/benchmark_inference.py --mode vllm --url http://localhost:$PORT_70B/v1 -n 10

echo ""
echo "--- 8B Sequential (20 requests) ---"
python scripts/benchmark_inference.py --mode vllm --url http://localhost:$PORT_8B/v1 -n 20

echo ""
echo "============================================"
echo "Test 2: Full Conversation Benchmark"
echo "============================================"
echo "Running 10 conversations with 70B user simulator + 8B agent..."
python scripts/benchmark_inference.py \
    --mode conversation \
    --url-70b http://localhost:$PORT_70B/v1 \
    --url-8b http://localhost:$PORT_8B/v1 \
    -n 10

# Cleanup
echo ""
echo "Cleaning up..."
kill $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null
wait $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null

echo ""
echo "============================================"
echo "BENCHMARK COMPLETE!"
echo "============================================"
echo ""
echo "Key metrics to compare with paper:"
echo "  - Paper: 2000 conversations/hour on H100x8"
echo "  - Expected A100x2 with 70B AWQ + 8B: ~100-300 conv/hr"
echo "  - Our old code: ~20 conv/hr"
echo ""
echo "If throughput is good, update experiment code to use vLLM."
echo ""
date