deployment issue in trt fp16

[sipie800]

I'm exporting a mask rcnn with convnext v2(mmdet and mmpretrain) to trt. However, while the fp32 model works well, fp16 model detect nothing. During exporting, tensorrt did complain a lot underflow of weight values.
We remove GRN and the issue is gone.

And when use maskrcnn along with convnext v1, nothing is wrong with both fp32 and fp16.

IMO despite GRN enhances feature output in a torch fp32 state, it does count on high numerical accuracy with that.
So if a fp16 trt - which is usually 1/4 flop than a fp32 trt for a cnn - can't be used for a convnext v2, its speed advantage over vit may not be true actually.