#!/bin/bash
#SBATCH --job-name=ctx_forget
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuH200x8-interactive
#SBATCH --gres=gpu:4
#SBATCH --time=01:00:00
#SBATCH --mem=200G
#SBATCH --cpus-per-task=32
#SBATCH --output=%x-%j.out
#SBATCH --error=%x-%j.err

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

source ~/miniforge3/etc/profile.d/conda.sh
conda activate eval

export HF_HOME=/projects/bfqt/users/yurenh2/hf_cache/huggingface
export TRANSFORMERS_CACHE=/projects/bfqt/users/yurenh2/hf_cache/huggingface
export PYTHONPATH=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model:$PYTHONPATH

PROFILE_PATH="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 "=== Starting vLLM servers ==="
date

# User simulator on GPUs 0,1 (70B)
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 (8B)
CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
    --model $AGENT_MODEL \
    --port 8003 --tensor-parallel-size 2 --gpu-memory-utilization 0.90 \
    --max-model-len 16384 --dtype bfloat16 &

# Wait for servers
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 (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

# Run contextual with reduced memory limits (4000 tokens, 15 turns)
CUDA_VISIBLE_DEVICES=2,3 python collaborativeagents/scripts/run_experiments.py \
    --methods contextual \
    --n-profiles 5 \
    --n-sessions 15 \
    --output-dir results/scale_test_contextual_forget \
    --profile-path $PROFILE_PATH \
    --datasets math-hard \
    --use-vllm --parallel-profiles 30 --no-batch-processing

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