#!/bin/bash
#SBATCH --job-name=run_expts_a100
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuA100x4
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=16
#SBATCH --gres=gpu:nvidia_a100:4
#SBATCH --mem=200G
#SBATCH --time=48:00:00
#SBATCH --output=logs/run_expts_a100_%j.out
#SBATCH --error=logs/run_expts_a100_%j.err

# Run experiments on 4x A100 80GB
# - 70B judge model with TP=4 (~140GB)
# - 8B PersonalizedLLM models (~16GB shared)

set -e

cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model

# Create logs and results directories
mkdir -p collaborativeagents/slurm/logs
mkdir -p collaborativeagents/results

echo "Starting experiments at $(date)"
echo "Job ID: $SLURM_JOB_ID"
echo "Node: $SLURMD_NODENAME"
echo "GPUs: $CUDA_VISIBLE_DEVICES"

# Activate environment
source /u/yurenh2/miniforge3/etc/profile.d/conda.sh
conda activate eval

# Check GPU availability
nvidia-smi

# Redirect HF cache to project space (avoid home quota issues)
export HF_HOME=/projects/bfqt/users/yurenh2/.cache/huggingface
mkdir -p $HF_HOME

# Add project to path
export PYTHONPATH="${PWD}/src:${PWD}/collaborativeagents:${PYTHONPATH}"

# Run experiments
cd collaborativeagents/scripts

# Full benchmark run
echo "Running full experiments..."
python run_experiments.py \
    --methods vanilla,all_memory,rag,rag_vector \
    --datasets math-500 \
    --n-profiles 20 \
    --n-sessions 5 \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl \
    --output-dir ../results/full_$(date +%Y%m%d_%H%M%S)

echo "Experiments completed at $(date)"