Brian Park

I have the same issue when doing densenet121 from torchvision's pretrained models.

@tcapelle I tried the LLaMA example on my M1 Pro 32GB. It's indeed slow, and I think that's mostly due to the weights being FP32. I haven't checked Mistral example...

@rovo79 Correct. As per #18, ANE API is closed source and not publicly accessible. I believe the only way to touch ANE today is via CoreML.

Hi Raidell, Thank you for your praise and suggestions. The graphs and analysis you show in the link are definitely interesting! I stopped developing and updating this repository a while...

Has there been an attempt or discussion to port in BNNS conv into MLX? [It's listed as TODO](https://github.com/ml-explore/mlx/blob/e6fecbb3e1e2ecb51247280f3738d670241777db/mlx/backend/accelerate/conv.cpp#L17). I've looked into it personally, but I'm noticing some limitations with BNNS...