| # DinoV3 Vision Transformer Huge (INT8 Quantized) | |
| INT8 quantized version of `facebook/dinov3-vith16plus-pretrain-lvd1689m` using BitsAndBytes. | |
| ## Model Details | |
| - **Base Model**: DinoV3 Vision Transformer Huge (840M parameters) | |
| - **Quantization**: INT8 weight-only quantization via BitsAndBytes | |
| - **Size**: ~845MB (from ~1.7GB original) | |
| - **Compression**: ~2x size reduction | |
| - **Accuracy Loss**: <1% typical | |
| ## Usage | |
| ```python | |
| from transformers import AutoModel, BitsAndBytesConfig | |
| # Load the INT8 quantized model | |
| model = AutoModel.from_pretrained( | |
| "Omdano/INT8-H16P", | |
| trust_remote_code=True, | |
| quantization_config=BitsAndBytesConfig(load_in_8bit=True), | |
| device_map="auto" | |
| ) | |
| # Use for feature extraction or classification | |
| outputs = model(pixel_values=inputs) | |
| ``` | |
| ## Benefits | |
| - **2x smaller** than full precision model | |
| - **Faster inference** on GPU | |
| - **Same API** as original DinoV3 | |
| - **Minimal accuracy loss** (<1%) | |
| ## Requirements | |
| ```bash | |
| pip install transformers bitsandbytes torch | |
| ``` | |
| ## Original Model | |
| Based on [facebook/dinov3-vith16plus-pretrain-lvd1689m](https://huggingface.co/facebook/dinov3-vith16plus-pretrain-lvd1689m) | |
| ## License | |
| Apache 2.0 (same as original DinoV3) | |