[Feature] Add full precision context to kfiou loss

[fengjk12138]

What's the feature?

version

dev-1.x

situation

I use kfiou on my dataset. In order to speed up training, I try to use AmpOptimWrapper. There is an error

RuntimeError: “lu_cuda” not implemented for ‘Half’

The Pytorch Community is believed that this part of the operation should be full precision operation. RuntimeError: “lu_cuda” not implemented for ‘Half’

Feature request

Add full precision context to this part of the operation. Make it easier for future users to use.

Any other context?

No response