#!/bin/bash
#SBATCH --job-name=rag_test
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuH200x8-interactive
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --gres=gpu:4
#SBATCH --mem=200G
#SBATCH --time=00:40:00
#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/rag_test-%j.out
#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/rag_test-%j.err

# Test with:
# 1. Skip reranking when few candidates
# 2. Reduced vLLM memory (0.35 for agent) to leave room for reranker

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="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/src:$PYTHONPATH"

PROFILE_PATH="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents/data/complex_profiles_v2/profiles_200.jsonl"
AGENT_MODEL="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
USER_MODEL="meta-llama/Llama-3.1-70B-Instruct"

echo "=== RAG Test v4 ==="
echo "Changes: Skip rerank when <k candidates, reduced vLLM memory (0.35)"
echo "5 profiles, 15 sessions"
date

# Clear empty store
> /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store/memory_cards.jsonl
rm -f /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store/memory_embeddings.npy

# Start vLLM servers with adjusted memory
# User simulator: 0.90 (unchanged)
CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
    --model $USER_MODEL \
    --port 8004 --tensor-parallel-size 2 --gpu-memory-utilization 0.90 \
    --max-model-len 16384 --dtype bfloat16 --download-dir $HF_HOME &

# Agent: reduced from 0.45 to 0.35 to leave room for reranker/embedding
CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
    --model $AGENT_MODEL \
    --port 8003 --tensor-parallel-size 2 --gpu-memory-utilization 0.35 \
    --max-model-len 16384 --dtype bfloat16 &

echo "Waiting for vLLM servers..."
for i in {1..200}; do
    if curl -s http://localhost:8004/health > /dev/null 2>&1; then
        echo "User simulator ready after $((i*5))s"
        break
    fi
    sleep 5
done
for i in {1..60}; do
    if curl -s http://localhost:8003/health > /dev/null 2>&1; then
        echo "Agent ready after $((i*5))s"
        break
    fi
    sleep 5
done
sleep 5

OUTPUT_DIR="../results/rag_test_v4_$(date +%Y%m%d_%H%M%S)"

for METHOD in vanilla rag rag_vector; do
    echo ""
    echo "============================================"
    echo "Testing: $METHOD"
    echo "============================================"

    # Clear memory store before each method
    > /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store/memory_cards.jsonl
    rm -f /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/data/corpora/empty_store/memory_embeddings.npy

    date
    python scripts/run_experiments.py --methods $METHOD \
        --datasets math-hard --n-profiles 5 --n-sessions 15 --max-turns 15 \
        --use-vllm --no-batch-processing --parallel-profiles 5 \
        --output-dir $OUTPUT_DIR --profile-path $PROFILE_PATH

    echo "Method $METHOD completed"
done

echo ""
echo "=== Done ==="
date

pkill -f "vllm.entrypoints" 2>/dev/null || true