## 🐛 Bug <!-- A clear and concise description of what the bug is. --> ## To Reproduce <!-- It is really important for the team to have a quick repro, which requires no setup work. The quicker is the repro to be run, the higher the chances the bug will be addressed sooner. The best way to create quick repros is to create a Colab based on the following template: https://github.com/pytorch/xla/blob/master/TROUBLESHOOTING.md#using-debug_runpy-to-collect-debug-information Things to avoid in repros is the need to download datasets which require setting up keys or other login information, like Kaggle downloads for example. Another example are Colab which mount user's Google Drive storages. Using a fake data generator could be a solution, in case the dataset cannot be easily downloaded without setting up credentials: https://github.com/pytorch/xla/blob/784b4d4f21751a54be0029a95f47d3896561c2a9/test/test_train_mp_mnist.py#L65 --> ``` import torch import torchax torchax.enable_globally() A = torch.randn(8, requires_grad=True) A.sum().backward() ``` raises error.
🐛 Bug
To Reproduce
raises error.