gemma-tuner-multimodal

Why we included this project

Most Gemma fine-tuning pipelines assume you have a CUDA GPU sitting around. This one runs LoRA fine-tuning for text, image captioning or VQA, and audio instruction entirely on an Apple Silicon Mac via Metal Performance Shaders, so developers without an NVIDIA box can still adapt the model to their own data. The workflow stays compact: a wizard-style CLI, local CSV datasets, and optional streaming of training shards straight from GCS or BigQuery, which avoids copying terabytes onto your laptop first. A built-in browser visualizer shows the loss curve, attention patterns, and token-by-token predictions live during training, so you get the feedback a TensorBoard setup would give without the extra plumbing. For individual developers and small teams who want private, on-device fine-tunes of Gemma 4 and 3n, it's a practical way to avoid renting cloud GPUs.