I was doing everything PyTorch and then I switched to lightning to accomplish my goal easier, and you still have room for "low level" (with 1000 quote marks) development.

I think that implementing feature X, and advancing with my research/work are two different, maybe equally exciting tasks, and keeping them separate is more productive. If you dont, then you would end up implementing something similar to a lightning/mmcv/etc clone and hey, they already exist!