PyTorch warnings about .grad access - safe to ignore?

[classner]

Thanks for the amazing package! I just played around with it and found that PyTorch emits two warnings when using the package:

/path/to/python/lib/python3.10/site-packages/taichi/lang/kernel_impl.py:763: UserWarning: The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the .grad field to be populated for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations. (Triggered internally at aten/src/ATen/core/TensorBody.h:489.)
  if v.requires_grad and v.grad is None:
/path/to/python/lib/python3.10/site-packages/taichi_splatting/misc/autograd.py:8: UserWarning: The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the .grad field to be populated for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations. (Triggered internally at aten/src/ATen/core/TensorBody.h:489.)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              
  grads = [tensor.grad if tensor.grad is not None else None

It seems like gradients are still propagated, but those warning are a little concerning - are they safe to ignore? Can you provide more context on what's happening here? The first one is related to a Taichi kernel launch and looking at the source probably okay. The second one refers to taichi_splatting.misc.autograd and I'm not sure what the restore_grad function is trying to do (I can't find an active reference).

[oliver-batchelor]

These are annoying warnings that arise from taichi tinkering with the `.grad` value of a torch Tensor for it's autograd integration. There should be a better way for them to handle the gradient integration. I think they shouldn't touch the `.grad` value at all (for example, call it something else like `.taichi_grad` or allow tuples of `(value, grad)` as input to their kernel.grad function)! I have raised this with them here: taichi-dev/taichi#8339

…

On Mon, Mar 18, 2024 at 4:08 PM Christoph Lassner ***@***.***> wrote: Thanks for the amazing package! I just played around with it and found that PyTorch emits two warnings when using the package: /path/to/python/lib/python3.10/site-packages/taichi/lang/kernel_impl.py:763: UserWarning: The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the .grad field to be populated for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations. (Triggered internally at aten/src/ATen/core/TensorBody.h:489.) if v.requires_grad and v.grad is None: /path/to/python/lib/python3.10/site-packages/taichi_splatting/misc/autograd.py:8: UserWarning: The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the .grad field to be populated for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations. (Triggered internally at aten/src/ATen/core/TensorBody.h:489.) grads = [tensor.grad if tensor.grad is not None else None It seems like gradients are still propagated, but those warning are a little concerning - are they safe to ignore? Can you provide more context on what's happening here? The first one is related to a Taichi kernel launch and looking at the source *probably* okay. The second one refers to taichi_splatting.misc.autograd and I'm not sure what the restore_grad function is trying to do (I can't find an active reference). — Reply to this email directly, view it on GitHub <#8>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAITRZMLCAHFB5VQJ3BVXZLYYZLDRAVCNFSM6AAAAABE2ZAK5CVHI2DSMVQWIX3LMV43ASLTON2WKOZSGE4TCMJTGM2DKMQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[classner]

Thanks a lot for the quick response - the linked issue explains this in much more detail. Fingers crossed this will be fixed upstream in Taichi. :)