[kshyatt]

This should allow us to use some GPU specific paths which will avoid blocking the device

[lkdvos]

Definitely looks good to me, although I feel like there must be way more places where this is used. As a side note, I think all of these could be normalize! but AD doesn't like mutating, so should we introduce normalize!! in TensorKit for that, or simply wait for Enzyme to land so we can support that automatically?

[kshyatt]

I feel like there must be way more places where this is used

I grepped every instance of norm( and this is what I found 🤷

[kshyatt]

simply wait for Enzyme to land so we can support that automatically

This makes sense to me

[kshyatt]

CUDA failures here are the known ones from the stuff we need from TK.