Gemma3n example

[bil-ash]

Description of the bug:

Please add Gemma-3n-E2B-IT .tflite generation (not .task generation) along with quantization example.

Actual vs expected behavior:

No response

Any other information you'd like to share?

No response

[sitatec]

➕ 1

[ericrisco]

+1

[bil-ash]

Any updates?

[gaikwadrahul8]

Hi, @bil-ash
Thank you for your patience and I apologize for the delay in my response. Regarding your query, we've confirmed a current limitation that prevents the export of the Gemma-3n-E2B-it model as a single, fully functional .tflite file. This is due to the exporter lacking certain required kernel implementations related to MatFormer architecture.

We have logged this as a feature request for the relevant team to address. While we will provide an update as soon as we have new information, we are unable to offer a specific ETA at this time.

We appreciate your understanding and patience.

[bil-ash]

Any updates?

[shahizat]

hello @gaikwadrahul8, Can we convert the fine-tuned google/gemma-3n-E4B-it model to the .task format? If so, could you please tell me how to do it?

[luiscrvm]

@shahizat Did you check here https://huggingface.co/google/gemma-3n-E4B-it-litert-preview/tree/main? Under the "Files and Version"? The .task file is there (4.4GB)

[shahizat]

Hey @luiscrvm, thanks. If I fine-tuned my own google/gemma-3n-E4B-it model, which is in the safetensors format, how can I use it with the Google AI Edge Gallery Android app?

[gaikwadrahul8]

Hi, @bil-ash, @shahizat, @ericrisco, @sitatec, @luiscrvm
We sincerely appreciate the community's interest in converting Gemma-3n models for on-device use. We understand this is a critical workflow for many developers. The core problem is an architectural incompatibility between Gemma 3 and Gemma-3n. While ai-edge-torch provides a robust, direct conversion path for Gemma 3 it does not yet have native end-to-end support for Gemma-3n's new features. The Gemma-3n architecture is a significant departure from its predecessor introducing a new multimodal backbone with a MobileNet v5 vision encoder and a Universal Speech Model (USM) audio encoder. Furthermore, the language model core itself includes novel components like MatFormer and Per-Layer Embeddings (PLE).

These new non-standard layers are the reason you are encountering conversion errors. The ai-edge-torch library which relies on torch.export to trace the model's computational graph fails when it encounters operations for which there is no corresponding TFLite operator. This is a feature gap, not a bug.  

@shahizat for fine-tuned model given the current tooling limitations just to confirm, did you try PyTorch -> ONNX -> TFLite conversion path then re-assemble the converted .tflite components with the tokenizer into a single .task bundle. This can be done using the mediapipe.tasks.python.genai.bundler utility ?

We have logged this as a feature request for the relevant team to address. While we will provide an update as soon as we have new information, we are unable to offer a specific ETA at this time. We appreciate your understanding and patience.