webssl-mae700m-full2b-224 昇腾 NPU 部署文档

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

# 安装 PyTorch
pip install torch==2.9.0 --index-url https://download.pytorch.org/whl/cpu

# 安装 torch_npu
pip install torch_npu==2.9.0.post1

# 安装 transformers 及其他依赖
pip install transformers==4.57.6 Pillow numpy scipy

npu-smi info

model_files/
├── config.json
├── preprocessor_config.json
└── model.safetensors

# NPU FP32 推理
python inference.py \
  --model_path ./model_files \
  --device npu \
  --precision fp32 \
  --image path/to/your/image.jpg \
  --save_features \
  --output_dir ./outputs

# NPU FP16 推理
python inference.py \
  --model_path ./model_files \
  --device npu \
  --precision fp16 \
  --image path/to/your/image.jpg \
  --save_features \
  --output_dir ./outputs

python accuracy_run.py --precision fp32 --output_dir ./logs
python accuracy_run.py --precision fp16 --output_dir ./logs

python accuracy_run_perf.py --device npu --precision fp32 --output_dir ./logs
python accuracy_run_perf.py --device npu --precision fp16 --output_dir ./logs

python check_accuracy_run_perf.py --precision both --output_dir ./logs

============================================================
Web-SSL MAE ViT-H Inference on Ascend NPU
============================================================
Device selected: npu
NPU Device: Ascend910_9362
torch version: 2.9.0+cpu
torch_npu version: 2.9.0.post1+gitee7ba04

[1/4] Loading model and processor...
  Model loaded in 2.37s

[2/4] Preparing input data...
  Input shape: torch.Size([1, 3, 224, 224])

[3/4] Warm-up inference...

[4/4] Running benchmark inference...
  Output shape: torch.Size([1, 257, 1280])
  Output dtype: torch.float32
  Output mean: 0.010152
  Output std: 0.479154
  Output max: 6.903232
  Output min: -14.741646
  Average inference time (5 runs): 0.0146s
  NPU Memory: allocated=2478.5MB, reserved=2584.0MB
  Model load time: 2.37s
  Avg inference time: 0.0146s

Web-SSL MAE ViT-H NPU inference completed successfully!

[INFO] Using device: npu:0
[INFO] Using precision: fp32 (torch.float32)
[INFO] Loading model from ./model_files ...
[INFO] Model loaded. Parameters: ~700M
[INFO] No image provided, using random 224x224 image
[INFO] Input shape: torch.Size([1, 3, 224, 224])
[INFO] Warmup 3 rounds ...
[INFO] Running inference ...

===== Inference Result =====
Feature shape:        torch.Size([1, 257, 1280])
Feature dtype:        torch.float32
Feature mean:         0.009344
Feature std:          0.479671
Feature min:          -13.164553
Feature max:          7.015934
Inference latency:    15.00 ms
Throughput:           66.67 images/sec

[INFO] Features saved to ./outputs/features_npu_fp32.pt
[INFO] Metadata saved to ./outputs/meta_npu_fp32.json

[INFO] Inference completed successfully.

============================================================
Web-SSL MAE ViT-H CPU vs NPU Precision Verification
============================================================
NPU Device: Ascend910_9362
torch version: 2.9.0+cpu
torch_npu version: 2.9.0.post1+gitee7ba04

[1/5] Loading model and processor...
  Model loaded

[2/5] Preparing fixed input...
  Input shape: torch.Size([1, 3, 224, 224])

[3/5] Running CPU inference (baseline)...
  CPU output shape: (1, 257, 1280)
  CPU output mean: 0.010283, std: 0.479693

[4/5] Running NPU inference...
  NPU output shape: (1, 257, 1280)
  NPU output mean: 0.010280, std: 0.479144

[5/5] Comparing precision...
  Max absolute diff: 7.184887e-02
  Mean absolute diff: 1.831753e-03
  MSE: 6.355060e-06
  Relative L2 error: 5.254095e-03 (0.5254%)
  Cosine similarity: 0.99998784
  Max relative error (|cpu| > 1e-4): 4.881472e+01

  Primary metric: Relative L2 error = 0.5254% (threshold < 1%)
  Secondary metric: Cosine similarity = 0.99998784 (threshold > 0.999)

  ✅ Precision verification PASSED

Precision verification completed!

.
├── inference.py                 # 标准推理脚本
├── accuracy_run.py              # 精度评测脚本
├── accuracy_run_perf.py         # 性能评测脚本
├── check_accuracy_run_perf.py   # 验收检查脚本
├── webssl_npu_infer.py          # NPU推理与性能基准脚本
├── webssl_npu_verify.py         # CPU vs NPU精度验证脚本
├── webssl_npu_adaptation_report.md  # 完整适配报告
├── logs/                        # 评测日志与报告
│   ├── accuracy_report_fp32.json
│   ├── accuracy_report_fp16.json
│   ├── perf_report_npu:0_fp32.json
│   ├── perf_report_npu:0_fp16.json
│   └── run_*.log
├── screenshots/                 # 自验证截图
│   ├── inference_success.png
│   ├── accuracy_passed.png
│   └── perf_benchmark.png
├── model_files/                 # 模型权重（用户自备）
└── readme.md                    # 本部署文档

@article{fan2025scaling,
  title={Scaling Language-Free Visual Representation Learning},
  author={David Fan and Shengbang Tong and Jiachen Zhu and Koustuv Sinha and Zhuang Liu and Xinlei Chen and Michael Rabbat and Nicolas Ballas and Yann LeCun and Amir Bar and Saining Xie},
  year={2025},
  eprint={2504.01017},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}