Implement MLX-LM For Completely Local Fine Tunes

[martymcfly1]

Could the Apple MLX-LM framework be implemented so that a user could in theory complete a full local fine tune of an MLX model on their Mac?

[scosman]

Kiln prepares the data and runs the model when fine tuning, but doesn’t manage the tuning process. You need to bring your own tuning engine, a cloud one like fireworks or a local one like unsloth. AFAIK there isn’t a great option for MLX tuning now.

[martymcfly1]

I've had success using Kiln to create the fine tuning dataset then pass it through MLX-LM on a Mac (https://github.com/ml-explore/mlx-lm). As a manual workflow, it's not without its challenges but do-able. It could make a good feature to an already great project - completely private / offline fine tuning workflow within one solution.

[aniruddhaadak80]

Even if Kiln does not want to own the tuning engine itself, MLX-LM support could still be valuable if there is a clean adapter point for local Mac-based fine-tuning workflows. A lot of people want the data prep and evaluation structure from tools like Kiln while keeping the actual training loop under their own control.nnSo the interesting question to me is less should Kiln run MLX directly, and more whether Kiln can make external local tuners first class citizens instead of cloud-only companions.

[scosman]

It can work with any API, local or cloud.

The cloud folks tend to have simple APIs that take training datasets and a few params in well known format. The local ones tend to have lower level APIs that are super powerful, but require custom code/notebook per model (like unsloth).

Both have their advantages but they are different types of users. Exporting datasets is the current solve for more powerful tools like unsloth. You get kilns dataset management/creation, unsloths powerful code config.

Something like axolot might be the balance between them.