It does make sense, but what you want is currently not supported in PyTorch.
There is issue #7313 that discusses this.
I once implemented a little technical exploration (linked from there).
There still is some latent interest in doing this eventually, but sadly It would seem that it’s not a priority.
Apropos spectral_norm in the title of that issue: There you have a similar goal and can see that the current approach is - let’s say - somewhat elaborate.
Best regards
Thomas