jeffhataws

"import _XLAC" error usually happens when TensorFlow or JAX are installed in the same environment. @comaniac can you check?

Thanks @JackCaoG and @amithrm !

> This means no mixed precision at all will be used during training as this variable controls the autocast context manager. @sgugger could you help point me to the autocast...

> The autocast context manager is defined [here](https://github.com/huggingface/transformers/blob/f48d3314e42bf54accc9dd8fd8dc1bf4197b34c6/src/transformers/trainer.py#L2604). > > As for your question on `torch.autocast`, we can't use it as it's only in very recent versions of PyTorch and...

> Mmm we cannot patch torch like this in Transformers as it's too magical and might yield to hard-to-debug issues for the users. Thanks. Please take a look at the...

> Mmm we cannot patch torch like this in Transformers as it's too magical and might yield to hard-to-debug issues for the users. @sgugger looks like using cpu_amp did not...

> I don't understand why it is necessary to patch torch.cuda for something you are telling me will not use torch.cuda anyway. Looks like there is some specific neuroncore tests...

@sgugger I have posted a revert here https://github.com/huggingface/transformers/pull/22451 . Apologies for the extra work.

> @jeffhataws @JackCaoG can you trigger the CI? Starting.

@JackCaoG do you know why the build failed with "ERROR: Error initializing RemoteModule"?