MMYolo vs. MMDetection

[collinmccarthy]

Hello, can someone help me understand why MMYolo exists as a standalone project, rather than being (largely) incorporated into MMDetection?

From #10006, it sounds like the main differences are the fast training, and due to this, the license (Apache 2.0 license in MMDetection vs. GPL v3 license in MMYolo). If that was the only difference, I would expect some small amount of source code to be pulled out and licensed separately, but not an entire "project" like MMYolo. Specifically, this answer states:

Since some components of the fast training are derived from YOLOv5, these parts of the training process are under the GPL License, and cannot be merged into MMDetection.

I have a couple follow up questions.

1. Can anyone be more specific about what the "fast training" components are, and why they couldn't be incorporated into MMDetection, with the exception of the YOLOv5 parts?
2. Are there other key abstractions which make MMYolo more "future proof" than MMDetection? If so, why not incorporate these into MMDetection as well?

For context, I'm trying to use either MMYolo or MMDetection to experiment with RTMDet and RT-DETR for a research project (#10498) (although I don't know if this is going to be merged into MMDetection or MMYolo). I need to be able to reproduce the accuracy and inference speed from the RTMDet paper, but if training takes a bit longer that's okay. And if I could easily incorporate some techniques from MMYolo into an MMDetection plug in to get most of the speed up while maintaining the Apache 2.0 license that would be amazing too.

I have looked at the following related issues/discussions: #10361, #10557, #10006. Thank you.