#!/bin/bash
#SBATCH --job-name=fs_%x
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuH200x8
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --gres=gpu:4
#SBATCH --mem=200G
#SBATCH --time=30:00:00
#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/fs_%x-%j.out
#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/fs_%x-%j.err

# Usage: sbatch --job-name=vanilla fullscale_method.sbatch vanilla
METHOD=$1

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"
# Full-precision 70B for user simulator (H200 143GB/GPU can handle it with TP=2)
USER_MODEL="meta-llama/Llama-3.1-70B-Instruct"

# vLLM memory and parallel workers for methods needing preference extractor
# These methods need GPU memory for embedding/reranker/extractor models on GPUs 2,3
if [[ "$METHOD" == "all_memory" || "$METHOD" == "rag" || "$METHOD" == "rag_vector" ]]; then
    AGENT_MEM=0.40  # Leave 60% free for embedding/reranker/extractor
    PARALLEL_PROFILES=30  # With CUDA_VISIBLE_DEVICES=2,3, extractor uses correct GPUs
else
    AGENT_MEM=0.90
    PARALLEL_PROFILES=50
fi

echo "=== Starting vLLM servers ==="
echo "Method: $METHOD"
echo "User simulator: $USER_MODEL (70B full-precision)"
echo "Agent: $AGENT_MODEL (8B)"
echo "Agent memory: $AGENT_MEM"
date

# User simulator on GPUs 0,1 (70B full-precision, ~70GB/GPU with TP=2)
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 on GPUs 2,3
CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
    --model $AGENT_MODEL \
    --port 8003 --tensor-parallel-size 2 --gpu-memory-utilization $AGENT_MEM \
    --max-model-len 16384 --dtype bfloat16 &

# Wait for 70B model to load (takes 9-12 minutes)
echo "Waiting for vLLM servers to be ready (this may take 10-15 minutes for 70B)..."
for i in {1..200}; do
    if curl -s http://localhost:8004/health > /dev/null 2>&1; then
        echo "User simulator (8004) ready after $((i*5)) seconds"
        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 (8003) ready after $((i*5)) seconds"
        break
    fi
    sleep 5
done
echo "Both vLLM servers ready"
sleep 10

# Batch processing only for vanilla
if [[ "$METHOD" == "vanilla" ]]; then
    EXTRA_ARGS="--use-batch-processing --batch-size 100"
else
    EXTRA_ARGS="--no-batch-processing"
fi

echo "Parallel profiles: $PARALLEL_PROFILES"

# Run experiment with CUDA_VISIBLE_DEVICES=2,3 so preference extractor/embedding/reranker
# use GPUs 2,3 (which have more headroom) instead of GPUs 0,1 (saturated by 70B model)
CUDA_VISIBLE_DEVICES=2,3 python scripts/run_experiments.py --methods $METHOD \
    --datasets math-hard --n-profiles 200 --n-sessions 30 --max-turns 15 \
    --use-vllm $EXTRA_ARGS --parallel-profiles $PARALLEL_PROFILES \
    --output-dir ../results/fullscale --profile-path $PROFILE_PATH

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