Richard Zou

@pytorchbot merge -g

@pytorchbot merge -g

@pytorchbot merge -g

Hey @Mahnoo06, MPS support is a ["prototype" feature in PyTorch](https://pytorch.org/blog/pytorch-1.12-released/#prototype-introducing-accelerated-pytorch-training-on-mac) so we haven't actively tried to support it yet. That being said I know M1 machines are popular, so we'll...

When you say second order gradients -- do you want the hessian? You can think of the model as a R^n -> R^1 function. The first order gradients have shape...

> I think it is a widely used trick if directly getting the Hessian is too costy. Is there a widely used trick to get the diagonal of the Hessian?

Thanks for the discussion! We'll take a look at the resources mentioned

I am not sure what is wrong here, but it sounds very similar to autograd.Function not working. @albanD do you have some context or reference on how nn Module hooks...

Cool, thanks for the analysis Alban. We are back to fixing the autograd.Function problem

@AlphaBetaGamma96 what are you using backward hooks for? Is it just per-sample-grad computation? If so we have other workarounds for that