About the --auto-scale-lr option

[brotherhoon-code]

I don't exactly understand about --auto-scale-lr.
If I have a batch size of 4096 trained in paper,
and my machine only supports a batch size of 64,
If I use this option, will the lr schedule adjustment be done after 64 steps (4096//64)?
If I want to use this,
Do I need to call --auto-scale-lr when calling train.py
from cmd in addition to what I define in config.py?

To summarize my question.

1. what exactly is auto-scale-lr for?
2. do I need cmd and config.py to utilize auto-scale-lr?
   (Or does it run automatically if it's in config)
3. it using like gradient accumlation?