字节跳动开源RL框架verl：让大模型强化学习效率提升20倍，已支持671B参数模型训练

# 以Qwen3-8B为例的GRPO训练命令
python -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=/data/gsm8k/train.parquet \
    data.val_files=/data/gsm8k/test.parquet \
    data.train_batch_size=1024 \
    data.max_prompt_length=512 \
    data.max_response_length=1024 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    actor_rollout_ref.model.path=Qwen/Qwen3-8B \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    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_per_gpu=40 \
    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.actor.entropy_coeff=0 \
    actor_rollout_ref.actor.strategy=fsdp2 \
    actor_rollout_ref.model.enable_gradient_checkpointing=False \
    actor_rollout_ref.actor.fsdp_config.param_offload=True \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=40 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
    actor_rollout_ref.rollout.n=5 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=40 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    actor_rollout_ref.ref.strategy=fsdp2 \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.logger=['console','tensorboard'] \
    trainer.project_name='verl_grpo_example_gsm8k_record' \
    trainer.experiment_name='qwen2_7b_function_rm_re' \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=1 \
    trainer.save_freq=20 \
    trainer.test_freq=5 \
    trainer.total_epochs=15

# 基础依赖安装
pip install -r requirements.txt
# 如需CUDA加速
pip install -r requirements-cuda.txt