CoolFace
Modelpublic

aufklarer/FunctionGemma-270M-CoreML-INT8-FP32-Stateful

sourceHugging Facegemmaupdated 3mo agoView on Hugging Face
0likes8downloads
Model Card

FunctionGemma 270M CoreML Stateful INT8

CoreML ML Program export of `google/functiongemma-270m-it` for on-device function calling.

  • —INT8 linear weights, FP32 compute
  • —256-token K/V cache stored as CoreML MLState
  • —Fixed specializations for 128-token padded prefill and 1-token decoding
  • —Requires iOS 18 / macOS 15 or newer

The model uses FunctionGemma's compact function-call syntax. Parse and validate the generated call against an application-side allowlist before executing any tool.

Contents

  • —FunctionGemmaStatefulDecoder.mlpackage: CoreML model package
  • —config.json: runtime input and cache metadata
  • —tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer assets

Validation

The export has been verified with an end-to-end weather tool call and a second-pass user-facing response after the tool result.

Runtime

Use the FunctionGemmaCoreMLStateful runtime in this repository's export directory. It performs a padded 128-token prefill followed by one-token decode calls, preserving MLState between invocations.