Until now, anyone who wanted to work with Apple's useful machine learning framework MLX had to use a Mac with as much M-Chip power as possible. Fans of tools such as LM Studio, which can be used to run local LLMs, swear by it on macOS and use MLX-optimized versions of Llama, Qwen or Mistral, which are designed to make better use of Apple SoCs and their unified memory. If some developers have their way, MLX could now also spread to other platforms. The GitHub project behind MLX is currently working on a CUDA backend. This means that Nvidia-based systems could use MLX in the future.

However, the project is still in its infancy. If it reaches product maturity, it is conceivable that MLX-optimized models could be adapted comparatively easily for the more powerful Nvidia GPUs, for example as part of training. In addition, MLX models could at least theoretically spread more quickly, as CUDA has a much larger community behind it.

MLX is specifically optimized for Apple Silicon and its Metal GPU interface. In the case of Nvidia GPUs, there would then be a switch from Metal to CUDA (which stands for Compute Unified Device Architecture). MLX comes with bindings for Swift, C and C++. The main developer behind the CUDA backend with the user name zcbenz comes from Japan. According to him, Apple is financing parts of the work.

"There are two main reasons for a CUDA backend," says the developer. "Firstly, CUDA supports unified memory. (...) Secondly, Nvidia hardware is widely used in the scientific field and on large computing clusters ["Massive Computations", writes zcbenz in the original]." It would therefore be useful to be able to write and test code locally on your Mac and then run it on supercomputers. "That would be a good developer experience."

It remains to be seen how quickly the project will progress. It was launched in the spring. This week, zcbenz announced that the CUDA backend is still under development and has not been officially released." More information will be provided "when things are in better shape".

(bsc)