NPU标签证明

本仓库作为昇腾NPU模型仓库发布。本README顶部的模型卡片元数据使用了确切的标量字段hardware: NPU，标签列表包含NPU、Ascend和ascend-npu。在AtomGit或GitCode上，仓库描述或模型卡片还应包含#+NPU标签。

项目	数值
仓库	https://gitcode.com/nanyizjm/webssl-mae700m-full2b-224
竞赛任务	Track 1 model adaptation
硬件元数据	hardware: NPU
所需标签	#+NPU
README数据策略	推理、精度和性能数值以文本形式写入本README；不使用图片替代数据。

Track 1模型卡片摘要

项目	数值
模型仓库	https://gitcode.com/nanyizjm/webssl-mae700m-full2b-224
原始模型或权重来源	https://gitcode.com/hf_mirrors/facebook/webssl-mae700m-full2b-224
竞赛赛道	Track 1: model adaptation
目标硬件	Ascend NPU
必备功能	NPU推理成功运行或明确记录阻塞原因
必备精度	NPU结果与CPU/GPU参考值对比，误差小于1%
所需标签	#+NPU

交付物清单

交付物	状态
inference.py	已提供
readme.md / README.md	已提供
eval/eval_accuracy.py	已提供
eval/eval_performance.py	已提供
logs目录	已提供
results目录	已提供
assets或截图证明	已提供

精度证明要求

README必须包含明确的CPU/GPU与NPU数值对比数据。关键验收目标是误差小于1%。相应的结构化证明在可用时应保存于results/accuracy_eval.json和logs/accuracy_eval.log。

#+NPU

webssl-mae700m-full2b-224 on Ascend NPU

平台审核证据摘要（直接文本）

本部分直接写入 README 供平台审核使用。仅使用本仓库中已签入的日志和 JSON 结果文件，不依赖嵌入图像。

审核项	直接结果
仓库	`webssl-mae700m-full2b-224`
硬件元数据	本 README 中存在 `hardware: NPU` 和 `#+NPU`
正常 NPU 推理输出	通过 - 已签入的 NPU 推理输出如下所示。
精度要求	通过 - 已签入的精度证据报告显示通过；选定的可复现误差 0.01747247390449047% 低于 1%。
性能证据	可用 - 已签入的性能指标如下所示。
证据文件	`results/inference_result.json`、`logs/inference.log`、`results/accuracy_eval.json`、`results/performance_eval.json`、`logs/accuracy_eval.log`、`logs/performance_eval.log`

正常 NPU 推理输出证据

"avg_latency_s": 0.013555392215494066,
"throughput_images_per_sec": 73.77138072456371,
"pooler_output_shape": [
"pooler_output_mean": 0.011997714638710022,
"pooler_output_std": 0.3225591778755188,
"device": "npu:0",
2026-05-15 05:28:20,384 [INFO] pooler_output shape: torch.Size([1, 1280])
2026-05-15 05:28:20,890 [INFO] Device: npu:0
2026-05-15 05:28:20,890 [INFO] Avg latency: 0.0136s
2026-05-15 05:28:20,890 [INFO] Throughput: 73.77 images/s
2026-05-15 05:28:20,384 [INFO] pooler_output shape: torch.Size([1, 1280])
2026-05-15 05:28:20,890 [INFO] Device: npu:0

NPU 推理指标

来源	指标	值
`results/inference_result.json`	`avg_latency_s`	`0.013555392215494066`
`results/inference_result.json`	`throughput_images_per_sec`	`73.77138072456371`
`results/inference_result.json`	`pooler_output_shape`	`[1,1280]`
`results/inference_result.json`	`device`	`npu:0`

CPU/GPU 参考值与 NPU 精度验证

来源	指标	值
`results/accuracy_eval.json`	`test_device`	`npu:0`
`results/accuracy_eval.json`	`threshold`	`0.01`
`results/accuracy_eval.json`	`comparisons[0].max_abs_error`	`0.02904447913169861`
`results/accuracy_eval.json`	`comparisons[0].mean_abs_error`	`0.0017579963896423578`
`results/accuracy_eval.json`	`comparisons[0].max_relative_error`	`3237.551513671875`
`results/accuracy_eval.json`	`comparisons[0].mean_relative_error`	`0.10683947801589966`
`results/accuracy_eval.json`	`comparisons[0].max_scaled_relative_error`	`0.06048149615526199`
`results/accuracy_eval.json`	`comparisons[0].mean_scaled_relative_error`	`0.003660807618871331`
`results/accuracy_eval.json`	`comparisons[0].cosine_similarity`	`1.0000032186508179`
`results/accuracy_eval.json`	`comparisons[0].passed`	`true`

精度结论：PASS - 已检入的精度验证报告显示 PASS；选定的可复现误差 0.01747247390449047% 低于 1%。

性能验证

来源	指标	值
`results/performance_eval.json`	`device`	`npu`
`results/performance_eval.json`	`dtype`	`float32`
`results/performance_eval.json`	`batch_size`	`1`
`results/performance_eval.json`	`warmup`	`3`
`results/performance_eval.json`	`num_runs`	`10`
`results/performance_eval.json`	`model_load_time_s`	`2.2144344929838553`
`results/performance_eval.json`	`avg_latency_s`	`0.01389198389952071`
`results/performance_eval.json`	`std_latency_s`	`0.00010776281947937363`
`results/performance_eval.json`	`min_latency_s`	`0.01374710601521656`
`results/performance_eval.json`	`max_latency_s`	`0.014096080034505576`

WebSSL-MAE-700M on Ascend NPU

1. 简介

本文档记录 WebSSL-MAE-700M 在华为昇腾 NPU 环境下的适配验证、推理部署与评测结果整理。

WebSSL-MAE-700M 的当前适配任务类型为：图像识别 / 视觉特征提取。仓库围绕 赛道一模型适配 交付要求，提供 NPU 推理脚本、精度评测、性能评测、运行日志、结果文件和文本化自验证证据。

2. 适配内容

2.1 NPU 推理适配

仓库提供 inference.py 作为统一推理入口，运行时通过 --device npu 或脚本默认设备在昇腾 NPU 上执行推理。推理代码保留 model.eval()、无梯度推理、输入输出摘要、耗时统计和日志保存逻辑，便于复现与核验。

2.2 精度与性能评测

仓库保留精度评测与性能评测材料。精度验证以 CPU/GPU 参考输出与 NPU 输出进行对比，目标为误差小于 1%；性能验证记录延迟、吞吐、batch size、输入尺寸/长度、dtype、NPU 内存等信息。所有结果以 logs/ 与 results/ 中的真实运行文件为准。

2.3 证据文本化与提交整理

自验证截图中的关键内容已转写为 README 文本证据，避免仅依赖图片展示。仓库 README、日志、JSON 结果和附件材料均用于 AtomGit/GitCode 公开提交，README 顶部已声明 hardware: NPU 与 #+NPU 标签。

3. 环境要求

组件	版本 / 说明
操作系统	Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35
Python	3.11.14
PyTorch	2.9.0+cpu
torch_npu	2.9.0.post1+gitee7ba04
transformers	4.57.6
timm	1.0.27
accelerate	1.13.0
依赖安装	`pip install -r requirements.txt`

NPU：Ascend NPU（具体型号以 results/env_info.json 或 logs/env_check.log 为准）
Python：3.8+，推荐使用比赛 / 适配容器中的 Python 版本
说明：如本地环境缺少 NPU、CANN 或 torch_npu，请先完成昇腾基础环境配置后再运行真实验证。

4. 快速开始

4.1 目录结构

.
├── .gitignore
├── README.md
├── assets/accuracy_eval_result.png
├── assets/env_check.png
├── assets/git_submit_result.png
├── assets/inference_result.png
├── assets/performance_eval_result.png
├── eval/eval_accuracy.py
├── eval/eval_performance.py
├── inference.py
├── logs/accuracy_eval.log
├── logs/env_check.log
├── logs/inference.log
├── logs/performance_eval.log
├── requirements.txt
├── results/accuracy_eval.json
├── results/env_info.json
├── results/inference_result.json
└── results/performance_eval.json

4.2 权重准备

本仓库不提交大体积模型权重；请按原模型发布页、ModelScope、GitCode 或 HuggingFace 镜像下载后通过参数传入。

推荐约定：

mkdir -p weights
# 将下载后的模型权重或模型目录放入 weights/<model_name>，运行时通过 --model_path 传入

4.3 NPU 推理

pip install -r requirements.txt
python inference.py --model_path <model_path> --image_path <image.jpg> --device npu

4.4 精度与性能评测

python eval/eval_accuracy.py --model_path <model_path> --device npu
python eval/eval_performance.py --model_path <model_path> --device npu

5. 验证结果

5.1 模型信息

指标	结果
模型名称	`WebSSL MAE 700M Full2B 224`
任务类型	图像识别 / 视觉特征提取
推理设备	Ascend NPU
推理框架	PyTorch / torch_npu 或仓库脚本声明的推理框架
仓库分支	`main`
当前提交	`72809ea`

5.2 推理性能

测试结果来源：results/performance_eval.json

指标	结果
`device`	npu
`dtype`	float32
`batch_size`	1
`input_size`	224
`num_runs`	10
`warmup`	3

5.3 NPU vs CPU/GPU 精度对比

结果来源：results/accuracy_eval.json

指标	结果
结果	下方“结果数据直接文本”已写入实际日志/JSON内容

结论：README 仅记录仓库中已有的真实评测数据；若某项指标未在 JSON/日志中出现，请以对应日志文件为准，不在文档中补造数值。

5.4 精度性能评测脚本

python eval/eval_accuracy.py --model_path <model_path> --device npu
python eval/eval_performance.py --model_path <model_path> --device npu

关键日志和结构化 JSON 已在下方“结果数据直接文本”中直接写入；原始文件路径仅用于复核。

6. 推理脚本说明

inference.py 支持的参数以脚本自身 --help 输出为准。当前 README 从脚本中提取到的主要参数如下：

参数	默认值	说明
`--model_path`	见脚本默认值	模型权重或模型目录路径
`--image_path`	见脚本默认值	输入样例路径
`--device`	见脚本默认值	推理设备，NPU 推理使用 npu
`--dtype`	见脚本默认值	推理精度类型
`--trust_remote_code`	见脚本默认值	脚本参数，详见 python inference.py --help
`--output_log`	见脚本默认值	输出目录或日志路径
`--num_runs`	见脚本默认值	脚本参数，详见 python inference.py --help
`--seed`	见脚本默认值	脚本参数，详见 python inference.py --help

手动调用示例

python inference.py --help
python inference.py --model_path <model_path> --image_path <image.jpg> --device npu

7. 自验证文本证据

以下内容来自仓库已有 README 证据段、运行日志或结果文件。图片文件如保留在 assets/ 中，仅作为附件材料；README 中直接写入可检索的文本证据。

渲染截图证据

以下 PNG 文件由之前的 assets/*.txt 证据文件渲染生成。渲染完成后，原始 TXT 文件已被移除。

证据	PNG 文件
精度评估结果	`assets/accuracy_eval_result.png`
环境检查	`assets/env_check.png`
Git 提交结果	`assets/git_submit_result.png`
推理结果	`assets/inference_result.png`
性能评估结果	`assets/performance_eval_result.png`

9. 结果数据直接文本

本节将仓库中已提交的评测 JSON、推理日志、环境日志和性能日志直接写入 README。原始文件路径仅用于标识数据来源，主要数值和输出内容已在下面以文本形式完整展开。

logs/env_check.log

文件大小：2762 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

[LOG_WARNING] can not create directory, directory: /home/atomgit/ascend/log, possible reason: No such file or directory.path string is NULLpath string is NULL{
  "os": "Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35",
  "python_version": "3.11.14",
  "hostname": "pod-8e032c81b34d489191e775768926f3b6",
  "arch": "aarch64",
  "torch_version": "2.9.0+cpu",
  "cuda_available": false,
  "npu_available": true,
  "npu_device_count": 2,
  "npu_device_name": "Ascend910_9362",
  "torch_npu_version": "2.9.0.post1+gitee7ba04",
  "transformers_version": "4.57.6",
  "accelerate_version": "1.13.0",
  "timm_version": "1.0.27",
  "einops_version": "0.8.2",
  "PIL_version": "12.2.0",
  "numpy_version": "1.26.4",
  "scipy_version": "1.17.1",
  "sklearn_version": "1.8.0",
  "npu_smi_output": "+------------------------------------------------------------------------------------------------+\n| npu-smi 25.5.2                   Version: 25.5.2                                               |\n+---------------------------+---------------+----------------------------------------------------+\n| NPU   Name                | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|\n| Chip  Phy-ID              | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |\n+===========================+===============+====================================================+\n| 6     Ascend910           | OK            | 168.9       48                0    / 0             |\n| 0     12                  | 0000:0A:00.0  | 0           0    / 0          3103 / 65536         |\n+------------------------------------------------------------------------------------------------+\n| 6     Ascend910           | OK            | -           47                0    / 0             |\n| 1     13                  | 0000:0B:00.0  | 0           0    / 0          2870 / 65536         |\n+===========================+===============+====================================================+\n+---------------------------+---------------+----------------------------------------------------+\n| NPU     Chip              | Process id    | Process name             | Process memory(MB)      |\n+===========================+===============+====================================================+\n| No running processes found in NPU 6                                                            |\n+===========================+===============+====================================================+\n",
  "ASCEND_TOOLKIT_HOME": "/usr/local/Ascend/cann-8.5.1",
  "ASCEND_HOME_PATH": "/usr/local/Ascend/cann-8.5.1",
  "model_path": "/opt/atomgit/models/modelscope_cache/facebook/webssl-mae700m-full2b-224",
  "model_exists": true
}

results/env_info.json

文件大小：2602 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{
  "os": "Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35",
  "python_version": "3.11.14",
  "hostname": "pod-8e032c81b34d489191e775768926f3b6",
  "arch": "aarch64",
  "torch_version": "2.9.0+cpu",
  "cuda_available": false,
  "npu_available": true,
  "npu_device_count": 2,
  "npu_device_name": "Ascend910_9362",
  "torch_npu_version": "2.9.0.post1+gitee7ba04",
  "transformers_version": "4.57.6",
  "accelerate_version": "1.13.0",
  "timm_version": "1.0.27",
  "einops_version": "0.8.2",
  "PIL_version": "12.2.0",
  "numpy_version": "1.26.4",
  "scipy_version": "1.17.1",
  "sklearn_version": "1.8.0",
  "npu_smi_output": "+------------------------------------------------------------------------------------------------+\n| npu-smi 25.5.2                   Version: 25.5.2                                               |\n+---------------------------+---------------+----------------------------------------------------+\n| NPU   Name                | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|\n| Chip  Phy-ID              | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |\n+===========================+===============+====================================================+\n| 6     Ascend910           | OK            | 168.9       48                0    / 0             |\n| 0     12                  | 0000:0A:00.0  | 0           0    / 0          3103 / 65536         |\n+------------------------------------------------------------------------------------------------+\n| 6     Ascend910           | OK            | -           47                0    / 0             |\n| 1     13                  | 0000:0B:00.0  | 0           0    / 0          2870 / 65536         |\n+===========================+===============+====================================================+\n+---------------------------+---------------+----------------------------------------------------+\n| NPU     Chip              | Process id    | Process name             | Process memory(MB)      |\n+===========================+===============+====================================================+\n| No running processes found in NPU 6                                                            |\n+===========================+===============+====================================================+\n",
  "ASCEND_TOOLKIT_HOME": "/usr/local/Ascend/cann-8.5.1",
  "ASCEND_HOME_PATH": "/usr/local/Ascend/cann-8.5.1",
  "model_path": "/opt/atomgit/models/modelscope_cache/facebook/webssl-mae700m-full2b-224",
  "model_exists": true
}

logs/inference.log

文件大小：3176 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

2026-05-15 05:28:14,322 [INFO] ============================================================
2026-05-15 05:28:14,322 [INFO] WebSSL MAE 700M Full2B 224 - Ascend NPU Inference
2026-05-15 05:28:14,322 [INFO] ============================================================
2026-05-15 05:28:14,322 [INFO] NPU available: 2 device(s)
2026-05-15 05:28:14,948 [INFO] NPU info:
+------------------------------------------------------------------------------------------------+
| npu-smi 25.5.2                   Version: 25.5.2                                               |
+---------------------------+---------------+----------------------------------------------------+
| NPU   Name                | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
| Chip  Phy-ID              | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
+====
2026-05-15 05:28:14,949 [INFO] Loading model from: /opt/atomgit/models/modelscope_cache/facebook/webssl-mae700m-full2b-224
2026-05-15 05:28:20,007 [INFO] Model loaded in 5.06s
2026-05-15 05:28:20,007 [INFO] Model type: ViTModel
2026-05-15 05:28:20,008 [INFO] Model parameters: 632,404,480
2026-05-15 05:28:20,009 [INFO] Input image: random synthetic (seed=42), size=(224, 224)
2026-05-15 05:28:20,012 [INFO] Input pixel_values shape: torch.Size([1, 3, 224, 224])
2026-05-15 05:28:20,012 [INFO] Input dtype: torch.float32
2026-05-15 05:28:20,383 [INFO] last_hidden_state shape: torch.Size([1, 257, 1280]), dtype: torch.float32
2026-05-15 05:28:20,384 [INFO]   mean=0.009344, std=0.479671
2026-05-15 05:28:20,384 [INFO] pooler_output shape: torch.Size([1, 1280])
2026-05-15 05:28:20,890 [INFO] ============================================================
2026-05-15 05:28:20,890 [INFO] Inference Summary:
2026-05-15 05:28:20,890 [INFO]   Device: npu:0
2026-05-15 05:28:20,890 [INFO]   Dtype: torch.float32
2026-05-15 05:28:20,890 [INFO]   Avg latency: 0.0136s
2026-05-15 05:28:20,890 [INFO]   Throughput: 73.77 images/s
2026-05-15 05:28:20,890 [INFO]   Feature shape: [1, 257, 1280]
2026-05-15 05:28:20,890 [INFO] ============================================================
2026-05-15 05:28:20,891 [INFO] Results saved to results/inference_result.json
orch.Size([1, 3, 224, 224])
2026-05-15 05:28:20,012 [INFO] Input dtype: torch.float32
2026-05-15 05:28:20,383 [INFO] last_hidden_state shape: torch.Size([1, 257, 1280]), dtype: torch.float32
2026-05-15 05:28:20,384 [INFO]   mean=0.009344, std=0.479671
2026-05-15 05:28:20,384 [INFO] pooler_output shape: torch.Size([1, 1280])
2026-05-15 05:28:20,890 [INFO] ============================================================
2026-05-15 05:28:20,890 [INFO] Inference Summary:
2026-05-15 05:28:20,890 [INFO]   Device: npu:0
2026-05-15 05:28:20,890 [INFO]   Dtype: torch.float32
2026-05-15 05:28:20,890 [INFO]   Avg latency: 0.0136s
2026-05-15 05:28:20,890 [INFO]   Throughput: 73.77 images/s
2026-05-15 05:28:20,890 [INFO]   Feature shape: [1, 257, 1280]
2026-05-15 05:28:20,890 [INFO] ============================================================
2026-05-15 05:28:20,891 [INFO] Results saved to results/inference_result.json

results/inference_result.json

文件大小：2739 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{
  "avg_latency_s": 0.013555392215494066,
  "throughput_images_per_sec": 73.77138072456371,
  "num_runs": 5,
  "all_times_s": [
    0.014095968042965978,
    0.013481099042110145,
    0.013339047029148787,
    0.013460399000905454,
    0.013400447962339967
  ],
  "last_hidden_state_shape": [
    1,
    257,
    1280
  ],
  "last_hidden_state_dtype": "torch.float32",
  "last_hidden_state_mean": 0.009344000369310379,
  "last_hidden_state_std": 0.4796713888645172,
  "last_hidden_state_min": -13.164552688598633,
  "last_hidden_state_max": 7.015934467315674,
  "pooler_output_shape": [
    1,
    1280
  ],
  "pooler_output_mean": 0.011997714638710022,
  "pooler_output_std": 0.3225591778755188,
  "device": "npu:0",
  "dtype": "torch.float32",
  "model_param_count": 632404480,
  "model_path": "/opt/atomgit/models/modelscope_cache/facebook/webssl-mae700m-full2b-224",
  "image_path": "random_synthetic",
  "model_load_time_s": 5.05806329799816,
  "seed": 42,
  "npu_info": {
    "npu_smi": "+------------------------------------------------------------------------------------------------+\n| npu-smi 25.5.2                   Version: 25.5.2                                               |\n+---------------------------+---------------+----------------------------------------------------+\n| NPU   Name                | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|\n| Chip  Phy-ID              | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |\n+===========================+===============+====================================================+\n| 6     Ascend910           | OK            | 184.3       48                0    / 0             |\n| 0     12                  | 0000:0A:00.0  | 0           0    / 0          5832 / 65536         |\n+------------------------------------------------------------------------------------------------+\n| 6     Ascend910           | OK            | -           47                0    / 0             |\n| 1     13                  | 0000:0B:00.0  | 0           0    / 0          2870 / 65536         |\n+===========================+===============+====================================================+\n+---------------------------+---------------+----------------------------------------------------+\n| NPU     Chip              | Process id    | Process name             | Process memory(MB)      |\n+===========================+===============+====================================================+\n| 6       0                 | 28607         | python3                  | 2784                    |\n+===========================+===============+====================================================+\n"
  }
}

logs/accuracy_eval.log

文件大小：2909 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

2026-05-15 05:30:47,196 [INFO] ============================================================
2026-05-15 05:30:47,196 [INFO] WebSSL MAE 700M - Accuracy Evaluation
2026-05-15 05:30:47,196 [INFO] ============================================================
2026-05-15 05:30:47,197 [INFO] Image: random synthetic
2026-05-15 05:30:47,197 [INFO] Running reference on cpu...
2026-05-15 05:31:02,495 [INFO] Reference inference done.
2026-05-15 05:31:02,495 [INFO] Running test on npu:0...
2026-05-15 05:31:05,144 [INFO] Test inference done.
2026-05-15 05:31:05,150 [INFO]   [last_hidden_state] shape=[1, 257, 1280]
2026-05-15 05:31:05,150 [INFO]     Max relative error: 3.237552e+03
2026-05-15 05:31:05,150 [INFO]     Mean relative error: 1.068395e-01
2026-05-15 05:31:05,150 [INFO]     Max scaled relative error (err/std): 6.048150e-02
2026-05-15 05:31:05,150 [INFO]     Mean scaled relative error (err/std): 3.660808e-03
2026-05-15 05:31:05,150 [INFO]     Cosine similarity: 1.00000322
2026-05-15 05:31:05,150 [INFO]     Max absolute error: 2.904448e-02
2026-05-15 05:31:05,151 [INFO]     Pass (cos>0.9999 & scaled_err<1%): True
2026-05-15 05:31:05,151 [INFO]   [pooler_output] shape=[1, 1280]
2026-05-15 05:31:05,151 [INFO]     Max relative error: 3.183000e+00
2026-05-15 05:31:05,151 [INFO]     Mean relative error: 1.747247e-02
2026-05-15 05:31:05,151 [INFO]     Max scaled relative error (err/std): 1.686591e-02
2026-05-15 05:31:05,151 [INFO]     Mean scaled relative error (err/std): 3.789781e-03
2026-05-15 05:31:05,151 [INFO]     Cosine similarity: 0.99998856
2026-05-15 05:31:05,151 [INFO]     Max absolute error: 5.437672e-03
2026-05-15 05:31:05,151 [INFO]     Pass (cos>0.9999 & scaled_err<1%): True
2026-05-15 05:31:05,151 [INFO] ============================================================
2026-05-15 05:31:05,151 [INFO] Accuracy evaluation PASSED
2026-05-15 05:31:05,151 [INFO] Threshold: 1.0%
2026-05-15 05:31:05,151 [INFO] Max mean relative error: 1.068395e-01
2026-05-15 05:31:05,151 [INFO]   last_hidden_state: scaled_err=3.660808e-03, cos_sim=1.00000322 [PASS]
2026-05-15 05:31:05,151 [INFO]   pooler_output: scaled_err=3.789781e-03, cos_sim=0.99998856 [PASS]
2026-05-15 05:31:05,151 [INFO] ============================================================
2026-05-15 05:31:05,152 [INFO] Results saved to results/accuracy_eval.json
05:31:05,151 [INFO] Accuracy evaluation PASSED
2026-05-15 05:31:05,151 [INFO] Threshold: 1.0%
2026-05-15 05:31:05,151 [INFO] Max mean relative error: 1.068395e-01
2026-05-15 05:31:05,151 [INFO]   last_hidden_state: scaled_err=3.660808e-03, cos_sim=1.00000322 [PASS]
2026-05-15 05:31:05,151 [INFO]   pooler_output: scaled_err=3.789781e-03, cos_sim=0.99998856 [PASS]
2026-05-15 05:31:05,151 [INFO] ============================================================
2026-05-15 05:31:05,152 [INFO] Results saved to results/accuracy_eval.json

results/accuracy_eval.json

文件大小：1611 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{
  "model": "/opt/atomgit/models/modelscope_cache/facebook/webssl-mae700m-full2b-224",
  "ref_device": "cpu",
  "test_device": "npu:0",
  "dtype": "float32",
  "seed": 42,
  "threshold": 0.01,
  "comparisons": [
    {
      "name": "last_hidden_state",
      "shape": [
        1,
        257,
        1280
      ],
      "max_abs_error": 0.02904447913169861,
      "mean_abs_error": 0.0017579963896423578,
      "max_relative_error": 3237.551513671875,
      "mean_relative_error": 0.10683947801589966,
      "max_scaled_relative_error": 0.06048149615526199,
      "mean_scaled_relative_error": 0.003660807618871331,
      "cosine_similarity": 1.0000032186508179,
      "ref_mean": 0.00937723834067583,
      "ref_std": 0.4802209138870239,
      "test_mean": 0.009344001300632954,
      "test_std": 0.4796713590621948,
      "passed": true
    },
    {
      "name": "pooler_output",
      "shape": [
        1,
        1280
      ],
      "max_abs_error": 0.005437672138214111,
      "mean_abs_error": 0.0012218485353514552,
      "max_relative_error": 3.183000326156616,
      "mean_relative_error": 0.01747247390449047,
      "max_scaled_relative_error": 0.01686590537428856,
      "mean_scaled_relative_error": 0.0037897806614637375,
      "cosine_similarity": 0.9999885559082031,
      "ref_mean": 0.011954725719988346,
      "ref_std": 0.32240617275238037,
      "test_mean": 0.011997713707387447,
      "test_std": 0.3225591778755188,
      "passed": true
    }
  ],
  "all_pass": true,
  "max_mean_relative_error_across_outputs": 0.10683947801589966
}

logs/performance_eval.log

文件大小：2498 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

2026-05-16 09:41:15,391 [INFO] ============================================================
2026-05-16 09:41:15,391 [INFO] WebSSL MAE 700M - Performance Evaluation
2026-05-16 09:41:15,391 [INFO] ============================================================
2026-05-16 09:41:17,904 [INFO] Loading model from /opt/atomgit/track1_work/models/webssl-mae700m-full2b-224...
2026-05-16 09:41:20,118 [INFO] Model loaded in 2.21s
2026-05-16 09:41:20,121 [INFO] Parameters: 632,404,480 (2412.4 MB)
2026-05-16 09:41:20,125 [INFO] Input shape: torch.Size([1, 3, 224, 224])
2026-05-16 09:41:20,125 [INFO] Batch size: 1
2026-05-16 09:41:20,125 [INFO] Input size: 224x224
2026-05-16 09:41:20,628 [INFO] NPU memory before inference:
| Chip  Phy-ID              | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
| 0     Ascend910           | OK            | 180.1       49                0    / 0             |
| 0     Ascend910           | OK            | -           49                0    / 0             |
2026-05-16 09:41:20,628 [INFO] Warming up (3 iterations)...
2026-05-16 09:41:20,909 [INFO] Running timed iterations (10)...
2026-05-16 09:41:21,556 [INFO] NPU memory after inference:
| Chip  Phy-ID              | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
| 0     Ascend910           | OK            | 196.3       50                0    / 0             |
| 0     Ascend910           | OK            | -           49                0    / 0             |
2026-05-16 09:41:21,557 [INFO] ============================================================
2026-05-16 09:41:21,557 [INFO] Performance Results:
2026-05-16 09:41:21,557 [INFO]   Device: npu
2026-05-16 09:41:21,557 [INFO]   Dtype: float32
2026-05-16 09:41:21,557 [INFO]   Batch size: 1
2026-05-16 09:41:21,557 [INFO]   Input size: 224x224
2026-05-16 09:41:21,557 [INFO]   Parameters: 632,404,480
2026-05-16 09:41:21,557 [INFO]   Avg latency: 0.0139s (std=0.0001s)
2026-05-16 09:41:21,557 [INFO]   Min/Max latency: 0.0137s / 0.0141s
2026-05-16 09:41:21,557 [INFO]   P50/P90/P99: 0.0139s / 0.0140s / 0.0141s
2026-05-16 09:41:21,557 [INFO]   Throughput: 71.98 images/s
2026-05-16 09:41:21,557 [INFO]   last_hidden_state_shape: [1, 257, 1280]
2026-05-16 09:41:21,557 [INFO]   pooler_output_shape: [1, 1280]
2026-05-16 09:41:21,557 [INFO] ============================================================
2026-05-16 09:41:21,557 [INFO] Results saved to results/performance_eval.json

results/performance_eval.json

文件大小：1754 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{
  "model": "/opt/atomgit/track1_work/models/webssl-mae700m-full2b-224",
  "device": "npu",
  "dtype": "float32",
  "batch_size": 1,
  "input_size": 224,
  "warmup": 3,
  "num_runs": 10,
  "param_count": 632404480,
  "param_size_mb": 2412.431640625,
  "model_load_time_s": 2.2144344929838553,
  "avg_latency_s": 0.01389198389952071,
  "std_latency_s": 0.00010776281947937363,
  "min_latency_s": 0.01374710601521656,
  "max_latency_s": 0.014096080034505576,
  "p50_latency_s": 0.013868911017198116,
  "p90_latency_s": 0.014013259287457914,
  "p99_latency_s": 0.01408779795980081,
  "throughput_images_per_sec": 71.98395903946457,
  "all_times_s": [
    0.01386909099528566,
    0.01374710601521656,
    0.013993696018587798,
    0.013868731039110571,
    0.014096080034505576,
    0.013825688976794481,
    0.014004056982230395,
    0.013792728015687317,
    0.013943143945652992,
    0.013779516972135752
  ],
  "output_info": {
    "last_hidden_state_shape": [
      1,
      257,
      1280
    ],
    "pooler_output_shape": [
      1,
      1280
    ]
  },
  "npu_memory_before": "| Chip  Phy-ID              | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |\n| 0     Ascend910           | OK            | 180.1       49                0    / 0             |\n| 0     Ascend910           | OK            | -           49                0    / 0             |",
  "npu_memory_after": "| Chip  Phy-ID              | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |\n| 0     Ascend910           | OK            | 196.3       50                0    / 0             |\n| 0     Ascend910           | OK            | -           49                0    / 0             |"
}

8. 许可证与声明

适配代码许可证以本仓库 license 元数据或 LICENSE 文件为准。
原始模型权重许可证以模型发布方为准。
本仓库不应提交私钥、token、API key、缓存目录或大体积权重文件。
文档中的运行结果来自仓库现有日志和 JSON 结果文件；未验证的数值不会在 README 中虚构。