#!/bin/bash
#SBATCH --job-name=full_exp_v2
#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/full_exp_v2_%j.out
#SBATCH --error=logs/full_exp_v2_%j.err

# Full scale experiment v2 - with fixes:
# 1. Use stable datasets (math-500, humaneval) - avoid problematic ones
# 2. Reduced scale first (20 profiles, 15 sessions) to verify stability
# 3. Clear CUDA cache between methods

set -e

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

mkdir -p collaborativeagents/slurm/logs
mkdir -p collaborativeagents/results

echo "Starting FULL SCALE v2 experiment at $(date)"
echo "Job ID: $SLURM_JOB_ID"
echo "Node: $SLURMD_NODENAME"
echo "GPUs: $CUDA_VISIBLE_DEVICES"

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

nvidia-smi

export HF_HOME=/projects/bfqt/users/yurenh2/hf_cache/huggingface
mkdir -p $HF_HOME

export PYTHONPATH="${PWD}/src:${PWD}/collaborativeagents:${PYTHONPATH}"

# Disable peer-to-peer memory access to avoid nvlink errors
export CUDA_VISIBLE_DEVICES=0,1,2,3
export NCCL_P2P_DISABLE=1

cd collaborativeagents/scripts

# FULL SCALE v2:
# - 20 profiles (reduced from 30 for stability)
# - 15 sessions per profile (enough to show learning)
# - 2 stable datasets: math-500, humaneval
# - All 7 methods
echo "Running FULL SCALE v2 experiments..."
python run_experiments.py \
    --methods vanilla,all_memory,rag,rag_vector,contextual,reflection,reflection_grpo \
    --datasets math-500,humaneval \
    --n-profiles 20 \
    --n-sessions 15 \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl \
    --output-dir ../results/full_experiment_v2_$(date +%Y%m%d_%H%M%S)

echo "Full experiment v2 completed at $(date)"