After commit 5a88c4b, I have observed a significant increase in compilation time for some warp kernels that previously compiled efficiently in my application. The increase seems to be related to changes in the vec/mat/quat autograd assignment, based on my bisection search of the commit history.

If needed, I would be happy to try to provide a minimal reproducible case by extracting a kernel from my application.

Thank you for any insights or suggestions!