| --- |
| library_name: aneforge |
| pipeline_tag: image-classification |
| tags: |
| - apple-neural-engine |
| - ane |
| - coreml-free |
| - on-device |
| - apple-silicon |
| - vit |
| - vision-transformer |
| - image-classification |
| license: mit |
| --- |
| |
| # Image classification on the Apple Neural Engine (via ANEForge) |
|
|
| [ANEForge](https://github.com/sbryngelson/ANEForge) runs computation on the Apple Neural |
| Engine (ANE) directly, without CoreML. `load_vit` loads a Hugging Face Vision Transformer |
| image classifier (`ViTForImageClassification`) from the Hub by repo id and runs the whole |
| forward pass on the engine. |
|
|
| This is a usage card, not a re-hosted model: it points at the upstream weights and shows how |
| to run them on the ANE. |
|
|
| ## Install |
|
|
| ```sh |
| pip install aneforge |
| ``` |
|
|
| Apple Silicon, macOS 14+. |
|
|
| ## Use |
|
|
| ```python |
| from aneforge.models import load_vit |
| from PIL import Image |
| |
| vit = load_vit("google/vit-base-patch16-224") # any HF ViT image classifier |
| image = Image.open("cat.jpg") |
| print(vit.classify(image, top_k=5)) # [(label, logit), ...]; forward on the ANE |
| # vit(image) -> raw logits [1, num_labels] |
| ``` |
|
|
| ## Measured |
|
|
| On an M5 Pro, `google/vit-base-patch16-224` runs the full forward in **~27 ms/image**, matching |
| the Hugging Face reference (same top-1, relerr 4e-3). Preprocessing uses the model's own |
| `AutoImageProcessor`. |
|
|
| ## Scope |
|
|
| ViT-family classifiers with a CLS token and a pre-norm encoder (`ViTForImageClassification` and |
| compatible DeiT/BEiT-style models); both the modern and legacy HF weight namings are handled. |
| ResNet / ConvNeXt loaders are tracked as follow-up issues in the repo. |
|
|
| ## Why the ANE |
|
|
| The ANE is the fixed-function accelerator on every recent Apple device. In production it is |
| reachable only through CoreML, which can silently fall back to CPU/GPU; ANEForge compiles the |
| classifier to a single ANE program and dispatches it through the same daemon and kernel-driver |
| stack Apple's own frameworks use. |
|
|
| ## Links |
|
|
| - Code: https://github.com/sbryngelson/ANEForge |
| - Package: https://pypi.org/project/aneforge/ |
| - Paper: https://arxiv.org/abs/2606.17090 |
|
|
| ## Cite |
|
|
| > Bryngelson, S. H. *ANEForge: Python for direct computation on the Apple Neural Engine.* arXiv:2606.17090 (2026). |
|
|
