verl NaN in multi-node training

NaN appears after one update (see the figure below) but the same training command with 1 node has no issue. Any suggestions?

python3 -m verl.trainer.main_ppo \
    data.train_files=$HOME/data/math/train.parquet \
    data.val_files=/$HOME/data/math/train.parquet \
    data.train_batch_size=1024 \
    data.val_batch_size=1312 \
    data.max_prompt_length=512 \
    data.max_response_length=1024 \
    data.return_raw_chat=True \
    actor_rollout_ref.model.path=Qwen/Qwen2.5-14B \
    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=16 \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.grad_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.model.enable_gradient_checkpointing=True \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=32 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=4 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    critic.optim.lr=1e-5 \
    critic.model.use_remove_padding=True \
    critic.model.path=Qwen/Qwen2.5-14B \
    critic.model.enable_gradient_checkpointing=True \
    critic.ppo_micro_batch_size_per_gpu=32 \
    critic.model.fsdp_config.param_offload=False \
    critic.model.fsdp_config.grad_offload=False \
    critic.model.fsdp_config.optimizer_offload=False \
    algorithm.kl_ctrl.kl_coef=0.001 \
    trainer.critic_warmup=0 \
    trainer.logger=['console','wandb'] \
    trainer.project_name='my_rl_exps' \
    trainer.experiment_name=test \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=2 \
    trainer.save_freq=50 \
    trainer.test_freq=10 \
    trainer.total_epochs=15