Mingzhi Hu

I have the same problem. The reason might be that pytorch-lightning calls optimizer.step(closure=closure, **kwargs) in the [precision_plugin](https://github.com/Lightning-AI/lightning/blob/1.6.4/pytorch_lightning/plugins/precision/precision_plugin.py#L155) and .grad is None before calling .backward.

> What's the bug? According to Keras API doc, y_true should be first and y_pred after. The current code may make the accuracy calculation wrong. https://keras.io/api/metrics/accuracy_metrics/

I tried enabling both ipex and bf16, but get different error message `the input and weight need have same data type`. Can you share the code so that i can...