DINOv3-ViT-B16 昇腾 NPU 部署指南

/data/
├── dinov3-vitb16-pretrain-lvd1689m/  # 模型权重
│   ├── model.safetensors
│   ├── config.json
│   ├── preprocessor_config.json
│   └── ...
└── dinov3-vitb16-pretrain-lvd1689m-ascend/  # 适配代码
    ├── inference.py                    # 推理脚本
    └── README.md                       # 本文档

docker rm -f test-dinov3

docker run -itd --privileged --name=test-dinov3 --net=host --shm-size=500g \
  --device=/dev/davinci0 \
  --device=/dev/davinci1 \
  --device=/dev/davinci2 \
  --device=/dev/davinci3 \
  --device=/dev/davinci4 \
  --device=/dev/davinci5 \
  --device=/dev/davinci6 \
  --device=/dev/davinci7 \
  --device=/dev/davinci_manager \
  --device=/dev/devmm_svm \
  --device=/dev/hisi_hdc \
  -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
  -v /usr/local/Ascend/add-ons/:/usr/local/Ascend/add-ons/ \
  -v /usr/local/sbin/:/usr/local/sbin/ \
  -v /var/log/npu/slog/:/var/log/npu/slog \
  -v /var/log/npu/profiling/:/var/log/npu/profiling \
  -v /var/log/npu/dump/:/var/log/npu/dump \
  -v /var/log/npu/:/usr/slog \
  -v /etc/hccn.conf:/etc/hccn.conf \
  -v /data:/data \
  quay.io/ascend/vllm-ascend:v0.13.0 \
  bash

docker exec test-dinov3 bash -c "pip3 install safetensors pillow -q -i https://repo.huaweicloud.com/repository/pypi/simple/"

docker exec test-dinov3 bash -c "cd /data/ysws/agentsp/dinov3-vitb16-pretrain-lvd1689m-ascend && python inference.py \
    --model_path /data/ysws/agentsp/dinov3-vitb16-pretrain-lvd1689m \
    --precision_test \
    2>&1 | tee /data/ysws/agentsp/dinov3-vitb16-pretrain-lvd1689m-ascend/log.txt"

docker exec test-dinov3 bash -c "cd /data/ysws/agentsp/dinov3-vitb16-pretrain-lvd1689m-ascend && python inference.py \
    --model_path /data/ysws/agentsp/dinov3-vitb16-pretrain-lvd1689m \
    --image_path /tmp/test_image.jpg \
    --device npu:0 \
    2>&1 | tee /data/ysws/agentsp/dinov3-vitb16-pretrain-lvd1689m-ascend/log.txt"

========================================================
Precision Comparison: CPU vs NPU
========================================================
Max errors: sum=9.16e-05, mean=1.19e-07, std=2.98e-08
PASS: NPU precision within thresholds
========================================================
PRECISION TEST PASSED
========================================================

2026-05-11 07:13:14,968 - INFO - DINOv3-ViT-B16 昇腾 NPU 推理
2026-05-11 07:13:17,549 - INFO - 模型已加载到设备: npu:0
2026-05-11 07:13:17,549 - INFO - 开始预热...
2026-05-11 07:13:17,860 - INFO - 预热完成
2026-05-11 07:13:17,860 - INFO - 开始推理...
2026-05-11 07:13:17,862 - INFO - 推理耗时: 0.0178s
2026-05-11 07:13:17,862 - INFO - 池化输出形状: torch.Size([1, 768])
2026-05-11 07:13:17,862 - INFO - 隐藏状态形状: torch.Size([1, 201, 768])
2026-05-11 07:13:17,862 - INFO - 推理成功完成!