| --- |
| license: other |
| license_name: apache-2.0-and-gemma |
| license_link: LICENSE.md |
| base_model: |
| - google/gemma-4-E4B-it |
| - google/embeddinggemma-300m |
| tags: |
| - coreml |
| - apple-neural-engine |
| - on-device |
| --- |
| |
| ## Gemma4E4B |
|
|
| gemma-4-E4B-it for the chunked engine of |
| [john-rocky/CoreML-LLM](https://github.com/john-rocky/CoreML-LLM): four decode |
| chunks plus batched prefill (N = 2048), context 4096. |
|
|
| This conversion differs from `mlboydaisuke/gemma-4-E4B-coreml` in three ways: |
|
|
| - **RoPE tables fixed.** `cos_full.npy` / `sin_full.npy` implement Gemma 4's |
| proportional RoPE for full attention: only the first 64 of the 256 |
| frequencies rotate. Upstream rotates all of them, so the model loses track of |
| anything more than ~512 tokens back. |
| - **Band mask for prefill.** Prefill applies the 512-token band mask to the |
| sliding-window layers. |
| - **Multifunction chunks.** Each `chunkN.mlmodelc` holds two functions, |
| `decode_q1` (the default) and `prefill`, which share one copy of the weights. |
| The `verify_qK` functions are not included. |
|
|
| The tokenizer adds no BOS token, so prepend `<bos>` yourself. |
|
|
| ## EmbeddingGemma |
|
|
| EmbeddingGemma-300M from |
| [erjigit17/embeddinggemma-300m-ane-coreml](https://huggingface.co/erjigit17/embeddinggemma-300m-ane-coreml), |
| compiled to `model.mlmodelc`. |
|
|
| - Inputs: `input_ids` and `attention_mask`, each int32 `[1, 128]`. |
| - Output: `embedding`, `[1, 768]`. |
|
|
| ## Licenses |
|
|
| - **Gemma4E4B:** Apache 2.0, the license of `google/gemma-4-E4B-it`. |
| - **EmbeddingGemma:** Gemma is provided under and subject to the Gemma Terms of |
| Use found at [ai.google.dev/gemma/terms](https://ai.google.dev/gemma/terms). |
| Use is also subject to the |
| [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy). |
| See `EmbeddingGemma/NOTICE`. |
|
|
