dousery/functiongemma-mobile-actions-litertlm

FunctionGemma Mobile Actions — LiteRT-LM Export

LiteRT-LM (TensorFlow Lite) format of the merged FunctionGemma mobile-actions model.
Use this for on-device / edge inference with the ai-edge-litert runtime.

What's inside

• mobile-actions_q8_ekv1024.litertlm — quantized LiteRT-LM model (Q8, kv_cache=1024)
• base_llm_metadata.textproto — metadata with BOS/EOS token IDs

Base model: dousery/functiongemma-mobile-actions (merged LoRA + base).

Intended use

• Function-calling for mobile actions (create calendar events, emails, contacts, maps, Wi‑Fi, flashlight, etc.)
• On-device / edge scenarios where a LiteRT-LM (.litertlm) is needed.

How to use (Python, ai-edge-litert)

pip install ai-edge-litert-nightly ai-edge-torch-nightly import litert from pathlib import Path

model_path = Path("mobile-actions_q8_ekv1024.litertlm")

Load LiteRT-LM

interp = litert.Interpreter.from_file(model_path)

Conversion details

• Source: dousery/functiongemma-mobile-actions (merged)
• Export: converter.convert_to_litert via ai_edge_torch.generative.utilities
• Quantization: quantize="dynamic_int8"
• KV cache: kv_cache_max_len=1024
• Prefill seq len: 256
• Export layout: kv_cache.KV_LAYOUT_TRANSPOSED
• Tokenizer model sourced from unsloth/functiongemma-270m-it (SentencePiece)

Citation

@misc{functiongemma-mobile-actions-litertlm,
  title={FunctionGemma Mobile Actions — LiteRT-LM Export},
  author={dousery},
  year={2025},
  howpublished={\url{https://huggingface.co/dousery/functiongemma-mobile-actions-litertlm}}
}