#!/bin/bash
export PYTHONPATH="/shared/storage-01/users/mehri2/verl:$PYTHONPATH"
set -x
HYDRA_FULL_ERROR=1

train_data="/shared/storage-01/users/mehri2/mem/collaborativeagents/training/grpo_verl/data/session_level_reflection_grpo_train.parquet"
model_path="/shared/storage-01/users/mehri2/LLaMA-Factory/saves/llama-3.1-8b-instruct/full/sft_session_level_reflection/checkpoint-628"
reward_fn_path="/shared/storage-01/users/mehri2/mem/collaborativeagents/training/grpo_verl/verl_reward_functions.py"

max_prompt_length=2048
max_response_length=1024
train_batch_size=8
n_generations=8
# Effective batch size is 64

python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files="$train_data" \
    data.val_files="$train_data" \
    data.train_batch_size=$train_batch_size \
    data.max_prompt_length=$max_prompt_length \
    data.max_response_length=$max_response_length \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    data.prompt_key=prompt \
    data.reward_fn_key=data_source \
    actor_rollout_ref.model.path=$model_path \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=8 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=8 \
    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=8 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
    actor_rollout_ref.rollout.n=$n_generations \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8 \
    actor_rollout_ref.ref.fsdp_config.model_dtype=bfloat16 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    actor_rollout_ref.rollout.temperature=0.9 \
    actor_rollout_ref.rollout.top_p=0.9 \
    custom_reward_function.path=$reward_fn_path \
    custom_reward_function.name=compute_score \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.val_before_train=False \
    trainer.logger='["console","wandb"]' \
    trainer.project_name='collaborative-agent-reflection-grpo' \
    trainer.experiment_name='llama3.1-8b-verl-grpo-v3' \
    trainer.n_gpus_per_node=4 \
    trainer.nnodes=1 \
    trainer.save_freq=50 \
    trainer.test_freq=100 \
    trainer.total_epochs=1 \
    trainer.default_local_dir=/shared/storage-01/users/mehri2/mem/collaborativeagents/training/grpo_verl/results/v3 $@