DINOv3 ViT-7B NPU 适配

pip install torch transformers pillow numpy
# torch_npu 和 CANN 驱动请参考华为昇腾官方文档安装

pip install atomgit
python3 -m atomgit download hf_mirrors/facebook/dinov3-vit7b16-pretrain-lvd1689m -d ./weight/dinov3-vit7b16-pretrain-lvd1689m

python inference.py --image /path/to/image.jpg

import torch
import torch_npu
from transformers import AutoImageProcessor, AutoModel
from PIL import Image

model_path = "./weight/dinov3-vit7b16-pretrain-lvd1689m"

processor = AutoImageProcessor.from_pretrained(model_path)
model = AutoModel.from_pretrained(model_path, torch_dtype=torch.float16)
model = model.to("npu")

image = Image.open("image.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
inputs = {k: v.to("npu") for k, v in inputs.items()}

with torch.no_grad():
    outputs = model(**inputs)

features = outputs.pooler_output  # [1, 4096]

.
├── inference.py              # NPU 推理脚本
├── readme.md                 # 本文件
├── evaluation/
│   ├── accuracy_eval.py      # 精度评测源代码
│   ├── accuracy_log.txt      # 精度评测运行日志
│   ├── perf_eval.py          # 性能评测源代码
│   └── perf_log.txt          # 性能评测运行日志
└── weight/
    └── dinov3-vit7b16-pretrain-lvd1689m/   # 模型权重

@misc{simeoni2025dinov3,
  title={{DINOv3}},
  author={Sim{\'e}oni, Oriane and Vo, Huy V. and Seitzer, Maximilian and others},
  year={2025},
  eprint={2508.10104},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
}