#!/bin/bash
# Quick test of parallel vLLM processing on A100x4-interactive

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 - using only 8B model for user sim to fit in A100
# (70B AWQ needs TP=2 which leaves only 2 GPUs for 8B)
MODEL_8B="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
PORT_USER=8004
PORT_AGENT=8003

echo "============================================"
echo "Quick Parallel vLLM Test (A100x4)"
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

# For A100 test, use 8B for both user and agent (to test parallelism)
# In production, user would be 70B AWQ with TP=2

# Start user simulator server (8B) on GPU 0-1
echo "Starting 8B user simulator server (GPU 0-1)..."
CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_8B \
    --port $PORT_USER \
    --tensor-parallel-size 2 \
    --gpu-memory-utilization 0.85 \
    --max-model-len 4096 \
    --disable-log-requests \
    --dtype bfloat16 &
SERVER_USER_PID=$!

# Start agent server (8B) on GPU 2-3
echo "Starting 8B agent server (GPU 2-3)..."
CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_8B \
    --port $PORT_AGENT \
    --tensor-parallel-size 2 \
    --gpu-memory-utilization 0.85 \
    --max-model-len 4096 \
    --disable-log-requests \
    --dtype bfloat16 &
SERVER_AGENT_PID=$!

echo "Waiting for servers..."

# Wait for servers (up to 5 minutes - A100 needs more time than H200)
for i in $(seq 1 100); do
    READY_USER=$(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)) seconds"
        break
    fi
    if [ $((i % 20)) -eq 0 ]; then
        echo "  Still waiting... user=$READY_USER, agent=$READY_AGENT ($((i*3))s)"
    fi
    sleep 3
done

# Check health
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"

cd scripts

echo ""
echo "============================================"
echo "Running throughput tests..."
echo "============================================"
echo "Note: Using 8B for both user and agent (parallelism test)"
echo ""

# Test 1: Sequential (1 profile, 3 sessions)
echo "--- Test 1: Sequential (1 profile, 3 sessions) ---"
START=$(date +%s)
python run_experiments.py \
    --methods vanilla \
    --datasets mmlu \
    --n-profiles 1 \
    --n-sessions 3 \
    --use-vllm \
    --vllm-user-url http://localhost:$PORT_USER/v1 \
    --vllm-agent-url http://localhost:$PORT_AGENT/v1 \
    --parallel-profiles 1 \
    --output-dir ../results/a100_test_1 \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl 2>&1 | tail -30
END=$(date +%s)
ELAPSED_1=$((END-START))
echo ""
echo "Time for 1 profile (3 sessions): ${ELAPSED_1} seconds"
echo "Throughput: ~$(echo "scale=1; 3 * 3600 / $ELAPSED_1" | bc) sessions/hr"

# Test 2: Parallel (4 profiles, 3 sessions each = 12 total)
echo ""
echo "--- Test 2: Parallel (4 profiles, 3 sessions each = 12 total) ---"
START=$(date +%s)
python run_experiments.py \
    --methods vanilla \
    --datasets mmlu \
    --n-profiles 4 \
    --n-sessions 3 \
    --use-vllm \
    --vllm-user-url http://localhost:$PORT_USER/v1 \
    --vllm-agent-url http://localhost:$PORT_AGENT/v1 \
    --parallel-profiles 4 \
    --output-dir ../results/a100_test_4 \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl 2>&1 | tail -30
END=$(date +%s)
ELAPSED_4=$((END-START))
echo ""
echo "Time for 4 profiles (12 sessions): ${ELAPSED_4} seconds"
echo "Throughput: ~$(echo "scale=1; 12 * 3600 / $ELAPSED_4" | bc) sessions/hr"

# Test 3: More parallel (8 profiles)
echo ""
echo "--- Test 3: Parallel (8 profiles, 3 sessions each = 24 total) ---"
START=$(date +%s)
python run_experiments.py \
    --methods vanilla \
    --datasets mmlu \
    --n-profiles 8 \
    --n-sessions 3 \
    --use-vllm \
    --vllm-user-url http://localhost:$PORT_USER/v1 \
    --vllm-agent-url http://localhost:$PORT_AGENT/v1 \
    --parallel-profiles 8 \
    --output-dir ../results/a100_test_8 \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl 2>&1 | tail -30
END=$(date +%s)
ELAPSED_8=$((END-START))
echo ""
echo "Time for 8 profiles (24 sessions): ${ELAPSED_8} seconds"
echo "Throughput: ~$(echo "scale=1; 24 * 3600 / $ELAPSED_8" | bc) sessions/hr"

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

echo ""
echo "============================================"
echo "RESULTS SUMMARY"
echo "============================================"
echo ""
echo "1 profile  (3 sessions):  ${ELAPSED_1}s  -> $(echo "scale=0; 3 * 3600 / $ELAPSED_1" | bc) sessions/hr"
echo "4 profiles (12 sessions): ${ELAPSED_4}s  -> $(echo "scale=0; 12 * 3600 / $ELAPSED_4" | bc) sessions/hr"
echo "8 profiles (24 sessions): ${ELAPSED_8}s  -> $(echo "scale=0; 24 * 3600 / $ELAPSED_8" | bc) sessions/hr"
echo ""
echo "Speedup 4x parallel: $(echo "scale=2; ($ELAPSED_1 * 4) / $ELAPSED_4" | bc)x"
echo "Speedup 8x parallel: $(echo "scale=2; ($ELAPSED_1 * 8) / $ELAPSED_8" | bc)x"
echo ""
date