VTP-Base-f16d64 on Ascend NPU

pip install torch torch_npu torchvision transformers omegaconf pillow

# ModelScope (推荐)
pip install modelscope
python3 -c "from modelscope import snapshot_download; snapshot_download('MiniMax/VTP-Base-f16d64')"

# HuggingFace
huggingface-cli download MiniMaxAI/VTP-Base-f16d64 --local-dir ./VTP-Base-f16d64

# 自编码器重建
python3 inference.py --model ./VTP-Base-f16d64 --mode ae

# CLIP 零样本分类
python3 inference.py --model ./VTP-Base-f16d64 --mode clip --image cat.jpg

# 特征提取
python3 inference.py --model ./VTP-Base-f16d64 --mode feature

# 全部模式
python3 inference.py --model ./VTP-Base-f16d64 --mode all

import torch, torch_npu
from vtp.models.vtp_hf import VTPModel

# 加载模型到 NPU
model = VTPModel.from_pretrained("./VTP-Base-f16d64", torch_dtype=torch.float16)
model = model.to("npu:0").eval()

# 自编码器验证
import torchvision.transforms as T
from PIL import Image
img = T.ToTensor()(Image.open("input.png")).unsqueeze(0).to("npu:0")

with torch.no_grad(), torch.autocast("npu", dtype=torch.float16):
    latents = model.get_reconstruction_latents(img)  # Encode
    recon = model.get_latents_decoded_images(latents) # Decode
print(f"Latents: {latents.shape}")  # [1, 64, 16, 16]
print(f"Recon: {recon.shape}")      # [1, 3, 256, 256]

python3 evaluate.py --model ./VTP-Base-f16d64 --name VTP-Base-f16d64 --output-dir ./eval