Wrong results when running inference with pytorch and MPS backend

[tonytorm]

just writing little code as the issue is very easy to recreate

device_gpu = torch.device("mps") model = bundle.get_model(with_star=False).to(device_gpu, non_blocking=False) with torch.inference_mode(): emission, _ = model(waveform.to(device_gpu))

this produces completely wrong emissions, the same code provides very precise results if ran on the cpu device (at more and less half the speed unfortunately)

I am gonna report the same issue on the pytorch github as I'm not capable of understanding where the the actual problem is located, loading the model to cpu or gpu provides same result (in terms of the model object data) but the inferences return 2 substantially different results.

[tonytorm]

Hey, I got news, the issue is within the components.py auxiliary script of the wav2vec2 model, specifically in the forward function, I'll link you to a similar issue that was raised in the pytorch git which led to a workaround too.

https://github.com/pytorch/pytorch/issues/98212

the problem appears to be related to non-contiguous tensors