Help: How did gradient_accumulation_steps in PPOConfig works?

[0verf1t-bit]

I was wander how Accelerator calculate batch_size when set batch_size mini_batch_size gradient_accumulation_steps in PPOConfig, but when I see in ds print_user_config json said that total batch size was equl to batch_size * num_process * gradient_accumulation_steps, instead of mini_batch_size * num_process * gradient_accumulation_steps.
This is a bit different from what I expected, and when I set the gradient_accumulation_steps in the accelerator config.yaml at the same time, the accumulation was the same with accelerator config.yaml, so what did PPOConfig gradient_accumulation_steps act during the training step? It's so wierd.