#!/bin/bash
# Compare batch processing vs parallel profile processing on A100x4
#
# Expected result: Batch should be significantly faster because:
# - Turn-synchronous: ALL conversations processed at same turn together
# - Maximizes vLLM continuous batching
# - Fewer total HTTP requests

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

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

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

echo "Waiting for servers (up to 5 min)..."

# Wait for servers
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"

echo ""
echo "============================================"
echo "Test 1: NEW Batch Processing (20 samples)"
echo "============================================"
echo "This batches ALL user requests together, then ALL agent requests."
echo ""

START=$(date +%s)
python agents/batch_vllm_agent.py \
    http://localhost:$PORT_USER/v1 \
    http://localhost:$PORT_AGENT/v1 \
    20
END=$(date +%s)
ELAPSED_BATCH=$((END-START))
echo ""
echo "Batch processing time: ${ELAPSED_BATCH} seconds"

echo ""
echo "============================================"
echo "Test 2: OLD Parallel Profile Processing (20 samples)"
echo "============================================"
echo "This runs 20 profiles in parallel, but each makes separate requests."
echo ""

cd scripts
START=$(date +%s)
python run_experiments.py \
    --methods vanilla \
    --datasets mmlu \
    --n-profiles 20 \
    --n-sessions 1 \
    --use-vllm \
    --vllm-user-url http://localhost:$PORT_USER/v1 \
    --vllm-agent-url http://localhost:$PORT_AGENT/v1 \
    --parallel-profiles 20 \
    --output-dir ../results/batch_compare_parallel \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl 2>&1 | tail -30
END=$(date +%s)
ELAPSED_PARALLEL=$((END-START))
echo ""
echo "Parallel profile processing time: ${ELAPSED_PARALLEL} seconds"

cd ..

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

echo ""
echo "============================================"
echo "RESULTS COMPARISON"
echo "============================================"
echo ""
echo "NEW Batch processing (20 conv):    ${ELAPSED_BATCH}s"
echo "OLD Parallel profiles (20 conv):   ${ELAPSED_PARALLEL}s"
echo ""
if [ $ELAPSED_BATCH -gt 0 ]; then
    SPEEDUP=$(echo "scale=2; $ELAPSED_PARALLEL / $ELAPSED_BATCH" | bc)
    echo "Speedup with batch processing: ${SPEEDUP}x"
fi
echo ""
echo "Expected: Batch should be 5-10x faster due to:"
echo "  - Turn-synchronous processing (all convs at same turn batched)"
echo "  - Fewer HTTP request overhead"
echo "  - Better vLLM continuous batching utilization"
echo ""
date