FunctionGemma 270M — LiteRT-LM (Android)

LiteRT-LM exports of google/functiongemma-270m-it for on-device function calling, plus a separately loadable Control LoRA adapter. The original stock model.litertlm remains unchanged.

Files

The LoRA-enabled base includes 128, 512, and 1024-token prefill graphs, a 1024-token KV cache, the tokenizer, and the FunctionGemma chat template. Base weights use dynamic INT8 quantization with FP32 activations for CPU/XNNPACK.

Control adapter

The adapter routes a compact offline phone-control surface:

call_contact, dial_number, find_contact, find_music, play_music, stop_music, set_volume, and list_capabilities.

It was trained as an attention-only rank-8 MLX LoRA on all 18 layers, targeting Q/K/V/O projections with scale 20. LiteRT-LM 0.14 CPU accepts ranks 16 and 32, so the export losslessly pads rank 8 to rank 16: the additional A columns and B rows are exactly zero. The adapter contains 144 external, 64-KiB-aligned FP32 tensor buffers and lora_rank=16 metadata.

Use the compact prompt format that matches training:

<start_of_turn>developer
You are a model that can do function calling with the following functions
Available functions: call_contact, dial_number, find_contact, find_music, play_music, set_volume, list_capabilities.
Music state: idle.<end_of_turn>
<start_of_turn>user
set volume to five<end_of_turn>
<start_of_turn>model

Add stop_music to the available-function line only while music is playing. The adapter is not intended for the full declaration prompt used by the stock bundle.

Android Kotlin 0.14

LiteRT-LM 0.14's Kotlin EngineConfig does not expose the engine LoRA-rank setting used by the C and Python APIs. Therefore Kotlin applications should use the adapter-required model-lora16-android.litertlm compatibility bundle and pass the adapter to each conversation:

val engine = Engine(
    EngineConfig(
        modelPath = "/data/local/tmp/model-lora16-android.litertlm",
        backend = Backend.CPU(),
    ),
)
engine.initialize()

val conversation = engine.createConversation(
    ConversationConfig(
        samplerConfig = SamplerConfig(topK = 1, topP = 1.0, temperature = 0.0),
        loraConfig = LoraConfig(
            loraPath = "/data/local/tmp/control-r4-rank16.tflite",
        ),
    ),
)

Greedy decoding is recommended for deterministic tool syntax. Released Kotlin 0.14.0 also lacks a per-conversation output-token cap; close each one-shot conversation after the command completes.

Validation

The release candidate uses training checkpoint step 400.

The Android engine loaded in 641 ms with an existing XNNPACK cache. Latency is generation-only after engine initialization. The 136-example held-out set contains action, no-action, and music-state cases. Dynamic INT8 loses routing accuracy relative to MLX, so applications should confirm high-impact actions such as dialing and should validate the model against their own command set.

Artifact integrity

The source model revision is google/functiongemma-270m-it@39eccb091651513a5dfb56892d3714c1b5b8276c.

Workflow contract

Function calls use:

<start_function_call>call:name{arg:<escape>value<escape>}<end_function_call>

The stock model supports ordinary FunctionGemma declarations and tool-response turns. The Control adapter is designed for one command and one tool call per fresh conversation; multi-step chaining and long multi-turn slot filling were not evaluated.

Links

• speech-android — Android SDK and Control demo
• speech-models — training and export tooling
• soniqo.audio