DINOv2-large on Ascend NPU

# 加载 CANN 环境
source /usr/local/Ascend/ascend-toolkit/set_env.sh

# 如需固定可见卡
export ASCEND_RT_VISIBLE_DEVICES=0

# 模型从 ModelScope 下载（避免 HuggingFace 连接超时）
pip install modelscope -i https://repo.huaweicloud.com/repository/pypi/simple/

python3 -c "
from modelscope.hub.snapshot_download import snapshot_download
snapshot_download('facebook/dinov2-large', cache_dir='./.cache/modelscope')
"

#!/usr/bin/env python3
import torch_npu
from torch_npu.contrib import transfer_to_npu  # 自动迁移：cuda -> npu

import torch
import numpy as np
from transformers import AutoImageProcessor, AutoModel
from PIL import Image
import requests

# 配置设备
device = torch.device("npu:0")

# 从本地 ModelScope 缓存加载
model_path = "./.cache/modelscope/facebook/dinov2-large"
processor = AutoImageProcessor.from_pretrained(model_path, use_fast=False)
model = AutoModel.from_pretrained(model_path, torch_dtype=torch.bfloat16)
model = model.to(device)
model.eval()

# 加载图片
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")

# 预处理并推理
inputs = processor(images=image, return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}

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

last_hidden_states = outputs.last_hidden_state
cls_token = last_hidden_states[:, 0, :]      # [1, 1024] 全局特征
patch_tokens = last_hidden_states[:, 1:, :]  # [1, 256, 1024] 局部特征

print(f"CLS token shape: {cls_token.shape}")
print(f"Patch tokens shape: {patch_tokens.shape}")

python3 dinov2_large_npu_inference.py