Awni Hannun

[Feature] Support pickling array for bfloat16

7

Leaving this here in case someone wants to pick it up.

Inconsistency in compile with kwargs

2

The fact that some of the following work but some don't seems inconsistent and unexpected. Filling this here mostly so I don't forget about it. ```python import mlx.core as mx...

Our reductions are quite naive and can be less accurate particularly in lower precision (`mx.float16`). NumPy and PyTorch (MPS) seem to use more sophisticated reductions. E.g. for NumPy: ``` For...

[BUG] in-place updating of array slice with broadcasting

1

This was a bug in MLX core that I think got inherited in MLX Swift. The following should work rather than throwing a broadcast error: ```swift import MLX let a...

Closes #1076

Closes #1066 This is kind of a nanobind issue but one that they have refused to fix in the past. - Basically a `np.ndarray` is implicitly convertible to a `std::complex`...

Jits: - Sort - Scan - Arange - Softmax - Reduce Down to 10MB.

JIT compile option for binary minimization

2

- Adds a build flag `MLX_METAL_JIT` to reduce the Metal library size by using runtime compilation. - Big refactor of unary, binary, ternary, copy, scatter, gather to allow JIT compilation...

[Feature] Expose something like custom VJP in Python

2

Consider allowing a function to have a custom VJP function attached to it.

[Feature] MLX backend

7

Any interest in doing an MLX back-end so we can run this efficiently on Apple silicon? [MLX docs](https://ml-explore.github.io/mlx/)