It’s nice to know that I’m not the only one having this issue, thanks for the links. Yes, let’s hope that this gets addressed by Apple. I have tested some other, simpler, models and they work fine, but the Metal results cannot be trusted right now.

I've updated the example code to use Tensorflow 2.7.0, just to keep it up to date. As you say, there is no change in results. As a note in case someone from Apple sees this, the model includes an Attention layer. Here's the model summary: input_1 (InputLayer) → conv1d (Conv1D) → conv1d_1 (Conv1D) → conv1d_2 (Conv1D) → conv1d_transpose (Conv1DTranspose) → conv1d_3 (Conv1D) → conv1d_transpose_1 (Conv1DTranspose) → flatten (Flatten) → conv1d_transpose_2 (Conv1DTranspose) → attention_layer (Attention) → conv1d_transpose_3 (Conv1DTranspose) → flatten_1 (Flatten) → activation (Activation) → dense (Dense)

Great to know that the issue is being addressed. Please keep us posted on any updates. Kind regards, Joseba

I've checked that, once I update to tensorflow 2.8.0 tensorflow provides the correct results. I've updated the repo sample to reflect this.