Is the NUM_ENV_STEPS_TRAINED logged incorrectly, if not how to interpret it compared to NUM_MODULE_STEPS_TRAINED?

RLlib Configure Algorithm, Training, Evaluation, Scaling
1

I realized that learners/__all_modules__/num_env_steps_trained has a strangely large value, multiple times higher than learners/__all_modules__/num_module_steps_trained.

For example:

I have a batch size and sample size of 2048, train for 20 epochs with a minibatch size of 128. The resulting learners log is:

__all_modules__: {
  num_module_steps_trained: 40960,
  num_env_steps_trained: 655360
}

num_module_steps_trained makes sense to me it is:
num_module_steps_trained = 2048 samples * 20 epochs = 128 minibatch_size * (2048/128 minibatch cycles) * 20 epochs.

However, num_env_steps_trained makes no sense to me - it is 16 times higher. It is calculated:
2048 samples * 20 epochs * (2048 / 128 minibatch cycles) = 2048 * 320 iterations total

I assume this is a bug from the logging:

batch.env_steps(), despite being a minibatch of size 128, returns the full batch_size 2048. So for the 320 iterations 2048 is logged and summed up to 655360.

Shouldn’t this be logged differently and elsewhere, or is there another angle how I could interpret this number?