main.sh

set -x
export VLLM_ATTENTION_BACKEND=XFORMERS
python3 verl/trainer/main_ppo.py \
    algorithm.adv_estimator=grpo \
    data.train_files=./data/train.parquet \
    data.val_files=./data/valid.parquet \
    data.train_batch_size=8 \
    data.val_batch_size=8 \
    data.max_prompt_length=4096 \
    data.max_response_length=4096 \
    actor_rollout_ref.model.path=./models_hub/Qwen2.5-Coder-7B \
    actor_rollout_ref.actor.optim.lr=3e-7 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=256 \
    actor_rollout_ref.actor.ppo_micro_batch_size=64 \
    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.grad_offload=True \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
    actor_rollout_ref.rollout.log_prob_micro_batch_size=160 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.temperature=1.0 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.9 \
    actor_rollout_ref.rollout.n=16 \
    actor_rollout_ref.ref.log_prob_micro_batch_size=160 \
    actor_rollout_ref.ref.fsdp_config.param_offload=False \
    algorithm.kl_ctrl.kl_coef=0.001 \
    trainer.critic_warmup=0 \
    trainer.logger=["console"] \
    trainer.project_name="CP-Zero" \
    trainer.experiment_name="Qwen-7B-Coder" \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=1 \
    trainer.default_local_dir=./checkpoints \
    trainer.default_hdfs_dir=null \
    trainer.save_freq=500 \
    trainer.test_freq=50 \
    trainer.total_epochs=5 2>&1 | tee verl.log