🚀 NeMo Framework Now Supports Google Gemma 3n: Efficient Multimodal Fine-tuning Made Simple

[nemo-automation-bot[bot]]

Authors: @HuiyingLi, @akoumpa, @snowmanwwg, @Sylendran95

Introduction

Gemma 3n is a generative AI model that takes inputs from a variety of modalities, including images and audio, and is optimized for efficient resource usage and fast inference on everyday devices. It introduces innovations such as Per-Layer Embedding parameter caching and the MatFormer architecture, which help reduce compute and memory demands, making it ideal for lightweight deployments. Some key highlights:

• Optimized architecture featuring MatFormer's nested transformers, Per-Layer Embeddings (PLE), and KV cache sharing. These enable sub-model extraction, reduced GPU memory usage, and faster prefill speeds.

• Multimodal capabilities with integrated image and audio encoders alongside the language model, enabling diverse tasks across modalities.
• Pretrained checkpoints are available under the Gemma 3n releases on Hugging Face.

Today, we are excited to announce that NeMo Automodel now supports Gemma 3n, making it easier than ever to load, train, and inference with Gemma 3n models.

---

⚡ Fine-tuning Gemma 3n with NeMo Automodel

NeMo Framework's Automodel path ("Nemo AutoModel") offers day-0 support for 🤗Hugging Face models via a unified interface to load and finetune models across modalities, abstracting away backend complexity. With Gemma 3n support:

• Load models with a single from_pretrained call
• Finetune models using full parameter training or PEFT(LoRA) with predefined recipes
• Accelerate training with kernel optimizations
• Leverage FSDP2/nvFSDP for efficient distributed training

Check out our tutorial on SFT and PEFT for both Gemma 3 and Gemma 3n models!

🔍 Observations

Training Dynamics

During the first hundred optimization steps we observed suspiciously large gradients that quickly stabilize. While the run remains numerically stable after this "warm-up," overall convergence still lags behind Gemma-3. We continue to investigate the source of this discrepancy.

Our preliminary benchmark on vision and audio capabilities show some gap between Gemma 3n and existing alternatives. We will follow up with more concrete results later.

✨ Conclusion

Gemma 3n brings impressive efficiency and opens up new possibilities for multimodal tasks on devices. With NeMo Automodel, getting started and fine-tuning these efficient models require only a few commands!

We look forward to seeing what you build with Gemma 3n and NeMo Automodel. Check out the documentation guide for a full walkthrough, and reach out on GitHub Discussions if you have questions.

🔗 References

Gemma 3n on Hugging Face

NeMo Automodel

NeMo Automodel Gemma 3 Guide