BlossomTuneLLM-MLX: Federated LLM Fine-Tuning with Flower, natively on Apple Silicon

[mrs83]

Hello mlx-lm community, I'm happy to share a project I've been working on that combines the power of mlx-lm with the world of federated learning:

GitHub Repo: BlossomTuneLLM-MLX

The core idea was to answer a simple question: how can we enable multiple Mac users to collaboratively fine-tune a language model without ever sharing their private data?

This is particularly relevant for creating specialized models in privacy-sensitive domains (like healthcare or legal tech) or for research that requires diverse, real-world data.

By combining mlx-lm with a federated learning framework, we can leverage the hardware people already own, reducing the reliance on expensive GPUs and promoting a more sustainable and accessible approach to language models.

This project is the MLX-native evolution of an earlier codebase for FlowerTune LLM, BlossomTuneLLM

How it Works: mlx-lm + Flower

• Federated Orchestration with Flower: Flower Framework handles all the federated learning logic. A central server (Superlink) coordinates the training rounds, client selection, and parameter aggregation.
• Client-Side Training with mlx-lm: Each participant in the network runs a Flower client (Supernode) on their Mac. In each round, the client:
  • Receives the global LoRA/DoRA adapter weights from the server.
  • Loads its local data partition.
  • Uses the mlx-lm programmatic API (mlx_lm.tuner.train) to perform LoRA/DoRA fine-tuning.
  • Sends only the updated adapter weights back to the server.

The server only ever sees the aggregated model updates, and private data never leaves the device (supernode).

Flower made it easy to run a full simulation (with a centralized HF dataset, partitioned using flower-datasets) on a single machine to test the whole process in action, fine-tuning SmolLM2-Instruct (135M).

All you need is a Mac with Apple Silicon. Or multiple ones :) we're just getting started and would love to get feedback, suggestions, and contributions from the MLX-LM community.

Thanks for reading!