huggingface/ratchet

Conv!

Opened this issue · 0 comments

We currently have a really poor convolution kernel.
Would be great to implement something like this: https://github.com/ml-explore/mlx/blob/main/mlx/backend/metal/conv.cpp#L176