main_reinforce.sh

set -x

MODEL_PATH=your_model_path
export VLLM_ATTENTION_BACKEND=XFORMERS
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export MASTER_PORT=123456
export WANDB_MODE=offline
export MLFLOW_TRACKING_URI=xxx
export MLFLOW_EXPERIMENT_NAME=xxx
export HYDRA_FULL_ERROR=1

python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=reinforce_plus_plus \
    data.train_files=data/xxx/train.parquet \
    data.val_files=data/xxx/val.parquet \
    data.train_batch_size=8 \
    data.val_batch_size=8 \
    data.max_prompt_length=1000 \
    data.max_response_length=100 \
    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=8 \
    actor_rollout_ref.actor.use_kl_loss=False \
    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=True \
    actor_rollout_ref.actor.fsdp_config.grad_offload=True \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
    actor_rollout_ref.rollout.log_prob_micro_batch_size=8 \
    actor_rollout_ref.rollout.name=hf \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.temperature=1.5 \
    actor_rollout_ref.rollout.n=8 \
    actor_rollout_ref.ref.log_prob_micro_batch_size=8 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    actor_rollout_ref.actor.clip_ratio=0.28 \
    trainer.critic_warmup=0 \
    trainer.logger=['wandb','tensorboard','mlflow'] \
    trainer.project_name='xxx' \
    trainer.experiment_name='xxx' \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=1 \
    trainer.default_local_dir=xxx \
    trainer.default_hdfs_dir=null \
    trainer.save_freq=300 \
    trainer.test_freq=10 \
    trainer.total_epochs=1 $@ 2>&1 | tee reinforce.log