Awni Hannun
Awni Hannun
@Goekdeniz-Guelmez I'm not convinced it makes sense to merge this given that it's pretty niche/new and also already in `mlx-optimizers`, see [here](https://stockeh.github.io/mlx-optimizers/build/html/_autosummary/mlx_optimizers.Muon.html). Anyone have any thoughts on that? If someone...
Maybe it's time to revisit adding this to core? Wdyt? Seems like the optimizer is getting pretty popular.
Yes let's re-open it!
I did a bit more simplification / nits. I think it is in good shape. What would be really nice is if we can confirm the implementation is correct with...
``` cmake .. -DMLX_BUILD_CPU=ON -DMLX_BUILD_SAFETENSORS=OFF ``` is currently broken... but rather than fix it, I think we can just land this.
@barronalex we can close this right?
I think the problem here is that we want to avoid dispatching differently based on the device. My preference would be to add a complex matmul and gemv for Metal...
The ops based dispatch is only three matmuls? Could you do the kernel implementation with just three like we do in the ops? > My guess is that BlockMMA is...
Indeed looks like a bug.
The extension version should technically be pinned at 18.1 / the next MLX release but I will keep it at 18.0 so our tests clear for the time being and...