#!/bin/bash
# Quick test of parallel vLLM processing on H200x8-interactive
# Simplified version for 1 hour time limit

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}"

# Configuration
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 "Quick Parallel vLLM Test (H200)"
echo "============================================"
date
echo "Node: $(hostname)"
nvidia-smi --query-gpu=index,name,memory.total --format=csv
echo ""

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

# Start 70B AWQ server on GPU 0-1 (TP=2)
echo "Starting 70B AWQ server (GPU 0-1, TP=2)..."
CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_70B_AWQ \
    --port $PORT_70B \
    --tensor-parallel-size 2 \
    --gpu-memory-utilization 0.90 \
    --max-model-len 4096 \
    --disable-log-requests \
    --quantization awq \
    --dtype float16 &
SERVER_70B_PID=$!

# Start 8B server on GPU 2
echo "Starting 8B server (GPU 2)..."
CUDA_VISIBLE_DEVICES=2 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 "Waiting for servers (up to 5 min)..."

# Wait for servers
for i in $(seq 1 100); do
    READY_70B=$(curl -s http://localhost:$PORT_70B/health > /dev/null 2>&1 && echo 1 || echo 0)
    READY_8B=$(curl -s http://localhost:$PORT_8B/health > /dev/null 2>&1 && echo 1 || echo 0)

    if [ "$READY_70B" = "1" ] && [ "$READY_8B" = "1" ]; then
        echo "Both servers ready after $((i*3)) seconds"
        break
    fi
    if [ $((i % 20)) -eq 0 ]; then
        echo "  Still waiting... 70B=$READY_70B, 8B=$READY_8B ($((i*3))s)"
    fi
    sleep 3
done

# Check health
if ! curl -s http://localhost:$PORT_70B/health > /dev/null; then
    echo "ERROR: 70B server not healthy"; kill $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null; exit 1
fi
if ! curl -s http://localhost:$PORT_8B/health > /dev/null; then
    echo "ERROR: 8B server not healthy"; kill $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null; exit 1
fi
echo "✓ Both servers healthy"

cd scripts

echo ""
echo "============================================"
echo "Running throughput tests..."
echo "============================================"

# Test 1: Sequential (1 profile, 2 sessions)
echo ""
echo "--- Test 1: Sequential (1 profile) ---"
START=$(date +%s)
python run_experiments.py \
    --methods vanilla \
    --datasets mmlu \
    --n-profiles 1 \
    --n-sessions 2 \
    --use-vllm \
    --vllm-user-url http://localhost:$PORT_70B/v1 \
    --vllm-agent-url http://localhost:$PORT_8B/v1 \
    --parallel-profiles 1 \
    --output-dir ../results/quick_test_1 \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl 2>&1 | tail -20
END=$(date +%s)
echo "Time: $((END-START)) seconds"

# Test 2: Parallel (4 profiles, 2 sessions each)
echo ""
echo "--- Test 2: Parallel (4 profiles) ---"
START=$(date +%s)
python run_experiments.py \
    --methods vanilla \
    --datasets mmlu \
    --n-profiles 4 \
    --n-sessions 2 \
    --use-vllm \
    --vllm-user-url http://localhost:$PORT_70B/v1 \
    --vllm-agent-url http://localhost:$PORT_8B/v1 \
    --parallel-profiles 4 \
    --output-dir ../results/quick_test_4 \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl 2>&1 | tail -20
END=$(date +%s)
echo "Time: $((END-START)) seconds"

# Test 3: Parallel (8 profiles, 2 sessions each)
echo ""
echo "--- Test 3: Parallel (8 profiles) ---"
START=$(date +%s)
python run_experiments.py \
    --methods vanilla \
    --datasets mmlu \
    --n-profiles 8 \
    --n-sessions 2 \
    --use-vllm \
    --vllm-user-url http://localhost:$PORT_70B/v1 \
    --vllm-agent-url http://localhost:$PORT_8B/v1 \
    --parallel-profiles 8 \
    --output-dir ../results/quick_test_8 \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl 2>&1 | tail -20
END=$(date +%s)
echo "Time: $((END-START)) seconds"

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

echo ""
echo "============================================"
echo "TEST COMPLETE!"
echo "============================================"
echo ""
echo "Summary: Compare timing above"
echo "  - Sequential (1 profile): baseline"
echo "  - Parallel (4 profiles): should be faster per profile"
echo "  - Parallel (8 profiles): should show more speedup"
echo ""
date