aoiandroid/translategemma-4b-it-coreml

Stateful TranslateGemma 4B IT (Core ML)

Stateful Core ML export of google/translategemma-4b-it with KV-cache states for incremental decoding on Apple platforms.

Included Files

• —StatefulTranslateGemma4BITFP16.mlpackage
• —StatefulTranslateGemma4BITInt8PerChannel.mlpackage
• —StatefulTranslateGemma4BITInt4PerChannel.mlpackage
• —convert_stateful_translategemma_coreml.py
• —NOTICE

Model Interface

Inputs:

• —inputIds: int32, shape (1, queryLength)
• —fullAttentionMask: float16, shape (1, 1, queryLength, endStep)
• —slidingAttentionMask: float16, shape (1, 1, queryLength, endStep)

States:

• —keyCache: float16, shape (layers, 1, kvHeads, maxContext, headDim)
• —valueCache: float16, same shape as keyCache

Output:

• —logits: float16

Conversion Notes

• —Conversion target: iOS 18+ (ct.target.iOS18)
• —Stateful export via Core ML states (ct.StateType)
• —Gemma3 mixed-attention export with explicit fullAttentionMask and slidingAttentionMask inputs
• —StatefulTranslateGemma4BITFP16.mlpackage: smallest stable FP16-focused artifact
• —StatefulTranslateGemma4BITInt8PerChannel.mlpackage: working balanced-size quantized variant
• —StatefulTranslateGemma4BITInt4PerChannel.mlpackage: smallest working quantized variant validated in short decode runs

Base Model and License

• —Base model: https://huggingface.co/google/translategemma-4b-it
• —Gemma terms: https://ai.google.dev/gemma/terms

This repository contains a converted derivative of Gemma model weights. Use is subject to Gemma license terms and policies.