#!/bin/bash
set -x

python3 -u main_ppo_custom_reward_server.py \
    algorithm.adv_estimator=grpo \
    data.train_files=$DATA_DIR/train.parquet \
    data.val_files=$DATA_DIR/test.parquet \
    data.train_batch_size=32 \
    data.val_batch_size=32 \
    data.max_prompt_length=512 \
    data.max_response_length=1024 \
    actor_rollout_ref.model.path=$BASE_MODEL \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=32 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=32 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.model.enable_gradient_checkpointing=True \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=32 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=$ROLLOUT_TP_SIZE \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
    actor_rollout_ref.rollout.n=8 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.kl_ctrl.kl_coef=0.001 \
    trainer.critic_warmup=0 \
    trainer.logger=['console'] \
    trainer.project_name='verl_chain_sum_grpo' \
    trainer.experiment_name=$EXPERIMENT_NAME \
    trainer.n_gpus_per_node=$N_GPUS \
    trainer.nnodes=1 \
    trainer.save_freq=100 \
    trainer.test_freq=100 \
    trainer.total_epochs=15 $@ 2>&1 | tee verl_output.log