Fix Automatic Multi-GPU Distribution for Gemma 4 Models
Hugging Face · Bug Fix · · minor
Briefing for: Engineering
What happened
Hugging Face released Transformers v5.5.3 to address a bug in the `device_map="auto"` feature specifically for Gemma 4 architectures. This patch ensures that model weights are correctly distributed across available GPU and CPU memory without manual intervention.
Why it matters
If you are working with Gemma 4 on multi-GPU setups, this update prevents initialization errors and incorrect device assignment. It stabilizes the developer experience for this specific model family by enabling reliable automated memory management.
What this enables
- If you use Gemma 4, you can now use the `auto` device mapping to handle model loading across multiple devices safely.
- If you were seeing device assignment errors during model initialization, update to v5.5.3 to stabilize your inference or fine-tuning pipelines.
Get personalized AI briefings for your role at Changecast →