Markus Enzweiler

I have noticed a significantly decreased throughput vs. the info in the original README. Both on a 16 MB M1 Macbook Pro. I am also seeing that without this PR...

@awni I did not record the "Previous" results. It is taken from the original ```README.md```. I have tried to reproduce it with earlier mlx versions (down to 0.0.2) but am...

Updated requirements.txt to require mlx>=0.0.9. Rebased and force-pushed.

Tested with mlx 0.0.9. Ready to be merged.

Adding another data point here (M1 16 GB, first gen MB Pro) with mlx 0.0.7, PyTorch 2.1.0 and the ```benchmarking``` branch from https://github.com/SarthakYadav/mlx-examples. ## MNIST num_layers = 4 hidden_dim =...

@SunnyBeike In my experiments with small (image) CNNs, see above, I am not seeing different GPU frequencies according to ```asitop```. For both torch and mlx, I am getting near 100%...

> I was able to repro using the code @menzHSE provided. If you pass `retain_graph=True` to the `eval()` method it suppresses the error, but it trains much slower, as expected....

Edit: Fixed in https://github.com/ml-explore/mlx/pull/409 There is one existing issue that I am not sure how to handle. When loading the model with ```strict=True``` I am getting a shape error for...

Awesome! I'll keep working on it and will wait for the next mlx release anyway. Regarding the batch norm issue: I tried but could not reproduce it with a minimal...

Verified that https://github.com/ml-explore/mlx/pull/409 fixed the loading in strict mode with mlx@1d90a76