Yukio Siraichi

@JackCaoG Is this ok to land? All the CI test are passing, now.

I will test for the regression described [here](https://github.com/pytorch/xla/pull/6624#issuecomment-1984717508) on the GPU machine I have access to.

@lsy323 Could you help me checking if the regression is gone?

Thanks, @vanbasten23.

@vanbasten23 Thanks. I will give it a go.

I don't think #6296 is wrong and should be reverted, since I believe it to be the best way to compare against inductor: instantiate the module in the original accelerator,...

> do we know why bias is fp16? Not really sure. I still have to investigate.

But, yes, this seems like PyTorch is doing something weird.

I've just confirmed that instantiating the model with XLA device solves the error. i.e. changing the line below with `str(self.benchmark_experiment.get_device())` https://github.com/pytorch/xla/blob/423bb0b295319a692ee21787edbff50d07361db7/benchmarks/torchbench_model.py#L233

Yes. I do think that's better. I was just confirming that that was the change that caused these models to break.