Current state of Gemma3n support

[wild vapor]

Hello, not sure if this is the right place to put this question, but Unsloth currently just supports single gpu inference and training for text and vision for Gemma 3n right?

There's no official data parallel multi-gpu inference/training no official Vision GRPO, no multi-gpu vision GRPO, and definitely no implementation of any of the above that can runn on a kaggle environment with 2 t4s right?

[thorn veldt]

You are correct: as of now, Unsloth officially supports single-GPU inference and training for text and vision with Gemma 3n. There is no official, fully documented support for data-parallel multi-GPU inference or training, nor for multi-GPU Vision GRPO (Group Relative Policy Optimization) with Gemma 3n. Vision GRPO support is under development, and while some community forks and experimental setups exist, there is no out-of-the-box, officially maintained solution for multi-GPU (including on Kaggle with 2 T4s) for these features yet. Official multi-GPU support is planned but not yet released for general use, and current documentation and notebooks focus on single-GPU workflows for Gemma 3n vision and text tasks. See Unsloth Docs, multi-GPU support status, and recent PRs/discussions.

Would you like more detail on unofficial workarounds or the current state of experimental features?

Sources:

• Unsloth Docs
• Unsloth Multi-GPU Support
• Unsloth Issue: multi-GPU training
• Unsloth PR: VLMs support for GRPO
• Unsloth Blog: Gemma 3n

💡 **Hint:** Mention @RunLLM in the thread for followups.

[ember sable]

You can try using accelerate if you’re familiar and know how to set it up

[wild vapor]

this would be for data parallel right?

[ember sable]

accelerate supports different multi gpu setups. Dataparallel is one

[wild vapor]

Are there any notebooks doing this? I heard there were some scripts for multi-gpu training, but not sure if there's anything for multi-gpu inference, esp considering VLLM may not have the vision mode up yet?