[Question] Clarification on feature routing logic in GDNeck2 forward pass (p4 vs n4)

[CoolinYu]

Hello team,

While studying the Gold-YOLO neck implementations, I noticed a structural difference in the forward pass between GDNeck2 and the other neck variants (GDNeck / RepGDNeck).

I would like to kindly request a clarification on whether this is an intentional design choice or a potential typo in the feature routing logic.

Details

In GDNeck and RepGDNeck, the bottom-up information flow sequentially aggregates features from n4 into n5:

## inject low-level global info to n5
n5_adjacent_info = self.LAF_n5(n4, c5_half) # Uses n4
n5 = self.Inject_n5(n5_adjacent_info, high_global_info[1])

However, in GDNeck2, the first argument passes p4 instead of n4:

## inject low-level global info to n5
n5_adjacent_info = self.LAF_n5(p4, c5_half) # Uses p4
n5 = self.Inject_n5(n5_adjacent_info, high_global_info[1])

What I want to confirm

Since I noticed that the official implementation applies different necks based on model scales (e.g., RepGDNeck for n/s, GDNeck for m, and GDNeck2 for l), I highly suspect this might be a specialized design for larger models.

Could you please help clarify:

1. Is this shortcutting behavior in GDNeck2 (bypassing n4 and reaching back to p4 for the n5 layer aggregation) an intentional architectural scaling strategy designed specifically to balance latency/accuracy or prevent feature degradation in Large models?
2. Or is it a minor typo that accidentally diverged from the baseline GDNeck routing?
   Thank you for your time and for sharing this great work!