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

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

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

export HF_HOME=/projects/bfqt/users/yurenh2/hf_cache/huggingface
export PYTHONPATH="${PWD}/src:${PWD}/collaborativeagents:${PYTHONPATH}"

# Install required packages (ensure they're in the conda env)
echo "Installing required packages..."
pip install --quiet json-repair tenacity

# Test: Verify imports work
echo "Testing imports..."
python3 -c "from json_repair import repair_json; from tenacity import retry; print('Imports OK')"

# Start judge model (70B) on GPUs 2,3
MODEL_70B="meta-llama/Llama-3.1-70B-Instruct"

echo "Starting judge model..."
CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_70B --port 8004 --tensor-parallel-size 2 \
    --gpu-memory-utilization 0.95 --max-model-len 8192 \
    --download-dir $HF_HOME --dtype bfloat16 --disable-log-requests &

# Wait for server
for i in $(seq 1 60); do
    curl -s http://localhost:8004/health > /dev/null 2>&1 && break
    sleep 3
done
echo "Judge model ready"

# Run GRPO with minimal steps for testing
echo "Starting GRPO test (10 steps only)..."
cd collaborativeagents/training/grpo_verl

python -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=${PWD}/data/session_level_reflection_grpo_train.parquet \
    data.val_files=${PWD}/data/session_level_reflection_grpo_train.parquet \
    data.train_batch_size=8 \
    data.max_prompt_length=2048 \
    data.max_response_length=1024 \
    data.filter_overlong_prompts=True \
    data.truncation=error \
    data.prompt_key=prompt \
    data.reward_fn_key=data_source \
    actor_rollout_ref.model.path=/work/nvme/bfqt/yurenh2/sft_checkpoints/checkpoint-200 \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=4 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.003 \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.actor.entropy_coeff=0 \
    actor_rollout_ref.model.enable_gradient_checkpointing=True \
    actor_rollout_ref.actor.fsdp_config.model_dtype=bfloat16 \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=2 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
    actor_rollout_ref.rollout.n=4 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=2 \
    actor_rollout_ref.ref.fsdp_config.model_dtype=bfloat16 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    actor_rollout_ref.rollout.temperature=0.9 \
    actor_rollout_ref.rollout.top_p=0.9 \
    custom_reward_function.path=${PWD}/verl_reward_functions.py \
    custom_reward_function.name=compute_score \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.val_before_train=False \
    trainer.logger='["console"]' \
    trainer.project_name=grpo-test \
    trainer.experiment_name=llama3.1-8b-grpo-test \
    trainer.n_gpus_per_node=2 \
    trainer.nnodes=1 \
    trainer.save_freq=100 \
    trainer.test_freq=100 \
    trainer.total_training_steps=10 \
    trainer.default_local_dir=/scratch/bfqt/yurenh2/grpo_test_outputs

echo "GRPO test complete!"
pkill -f "vllm.entrypoints" 2>/dev/null || true