#!/bin/bash
#SBATCH --job-name=vllm_only
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuA100x4
#SBATCH --nodes=1
#SBATCH --gpus-per-node=1
#SBATCH --time=00:45:00
#SBATCH --mem=64G
#SBATCH --output=slurm/logs/vllm_only_%j.out
#SBATCH --error=slurm/logs/vllm_only_%j.err

# Test vLLM inference speed ONLY (skip transformers which OOMs)
set -e

cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents

# Activate conda environment
source /u/yurenh2/miniforge3/etc/profile.d/conda.sh
conda activate eval

# Install vLLM if not already installed (requires GPU node for CUDA compilation)
if ! python -c "import vllm" 2>/dev/null; then
    echo "Installing vLLM..."
    pip install vllm --quiet
fi

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

MODEL_8B="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
PORT=8003

echo ""
echo "============================================"
echo "Starting vLLM Server for 8B Model"
echo "============================================"

# Start vLLM server with memory optimization
python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_8B \
    --port $PORT \
    --gpu-memory-utilization 0.85 \
    --max-model-len 4096 \
    --disable-log-requests \
    --dtype bfloat16 &
SERVER_PID=$!
echo "vLLM Server PID: $SERVER_PID"

# Wait for server to be ready
echo "Waiting for server to start..."
for i in {1..90}; do
    if curl -s http://localhost:$PORT/health > /dev/null 2>&1; then
        echo "Server ready after $((i*2)) seconds"
        break
    fi
    sleep 2
done

# Check if server is up
if ! curl -s http://localhost:$PORT/health > /dev/null 2>&1; then
    echo "ERROR: vLLM server failed to start"
    cat slurm/logs/vllm_only_${SLURM_JOB_ID}.err | tail -50
    kill $SERVER_PID 2>/dev/null || true
    exit 1
fi

# Get model info
echo ""
echo "=== vLLM Server Info ==="
curl -s http://localhost:$PORT/v1/models | python -m json.tool 2>/dev/null || echo "Could not get model info"

echo ""
echo "============================================"
echo "Test 1: vLLM Sequential (20 requests)"
echo "============================================"
python scripts/benchmark_inference.py --mode vllm --url http://localhost:$PORT/v1 -n 20

echo ""
echo "============================================"
echo "Test 2: vLLM Sequential (50 requests)"
echo "============================================"
python scripts/benchmark_inference.py --mode vllm --url http://localhost:$PORT/v1 -n 50

echo ""
echo "============================================"
echo "Test 3: vLLM Concurrent 4 workers (50 req)"
echo "============================================"
python scripts/benchmark_inference.py --mode vllm --url http://localhost:$PORT/v1 -n 50 --concurrent

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

echo ""
echo "============================================"
echo "BENCHMARK COMPLETE!"
echo "============================================"
echo ""
echo "Key metrics to compare with paper:"
echo "  - Paper: 2000 conversations/hour on H100x8"
echo "  - Expected A100x1: ~200-500 conv/hr"
echo "  - Our old code: ~20 conv/hr (100x slower)"
echo ""
echo "If vLLM shows good throughput, we need to update"
echo "our experiment code to use vLLM instead of transformers."
echo ""
date