Richard Zou

@hturki yes, that's unfortunately correct. This is still top of mind for us, but it is a bit difficult to implement. We've planned to implement it sometime in the next...

@hturki is there an example of a custom backward function you want to use with functorch?

Thanks for the example. There are two types of autograd.Function: - ones that are written using PyTorch operations - ones that call into custom written CUDA kernels (or other custom...

We've finally gotten to revisiting this. I'm curious to hear about other use cases for this feature so we know what to prioritize. Calling on some folks (but please feel...

Those test are actually expected to fail, they are there to remind us that we need to sync our testing database with PyTorch's testing database :/. I should make this...

Thanks for the issue, @agiachris! Is there a specific model or inputs/targets you're running the above with? If not we can try to make some dummy models/inputs and try to...

We haven't thought about how functorch composes with the torch.distributions package but yes we need to figure that out. More concretely, the problem here is a .all() call in distributions:...

The requirements we should consider are: - developer velocity (for functorch devs) - developer velocity (for pytorch/pytorch devs) - internal customers - limiting internal breakages - The ability to release...

Thanks for reporting, @lrast. I assume that you're using PyTorch 1.11 and functorch 0.1.*. This problem should be fixed in the pytorch nightly + functorch main branch. If you're interested...

@xmser thanks for the issue. I'm trying to run the repro above -- what is the value of `mini_batch` and what is the shape of `partial_xx_tensor`?