Any plans on the new Metal 4 tensor API integration？

[kemchenj]

The introduction of the Metal 4 native Tensor API (MTLTensor) seems like a major opportunity for performance optimization.

Given that MLX is purpose-built for Apple silicon, is the team currently planning to integrate this new API into the MLX backend?

I'm hoping this integration could lead to significant speed boosts and better hardware utilization on the newest Apple Silicon chips that fully support Metal 4.

Any insight on the roadmap would be appreciated.

[awni]

We're evaluating which parts of the Metal 4 API we can use that will result in efficiency gains for MLX. I don't have a roadmap to share with you on this.. but you can keep an eye out on the repo as things develop.