That is disappointing. I thought perhaps the lack of response was because they were working on a fix for the new 0.3 release version. Seems they weren’t.

Hi Frameworks Engineers, There are a number of similar reports that have gone unacknowledged by the engineering team for the past month. Evaluating Trained/Saved LSTM Model from another system gives wrong results on M1 GPU ; TensorFlow model predictions are incorrect on M1 GPU; and Training LSTM: Low Accuracy on M1 GPU vs CPU. This report has been acknowledged, but it is unclear as to whether it, like this report, is related: Inconsistent results when performing inference in CPU vs Metal. Just wondering if the framework engineers can assess these reports (they all have instructions to reproduce the issue), to see whether they relate to the same underlying problem? And if not, whether they are being investigated for resolution? Like many others, I am very excited to be using the new Apple Hardware (GPUs) for machine learning, but I've been unable to get any sensible results from using Apple's Metal pluggable device implementation.