Gemma 4 Multimodal Fine-Tuner for Apple Silicon

A developer built a local fine-tuning tool to run Whisper and later Gemma models on an M2 Ultra Mac Studio.

They created a streaming pipeline to feed 15,000 hours of audio from Google Cloud Storage during training to avoid local storage limits.

The project was updated to add Gemma 4 support and split Gemma from Whisper fine-tuning, but the developer faced frequent OOMs on 64GB RAM.

The author published the code for others to use, fork, and improve.

The tool exists partly because current platforms like MLX don't support audio fine-tuning for Gemma 4.