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,且标签列表包含 NPU、Ascend 和 ascend-npu。仓库描述或模型卡片在 AtomGit 或 GitCode 上还应包含 #+NPU 标签。
| 项目 | 数值 |
|---|---|
| 仓库 | https://gitcode.com/nanyizjm/phikon-v2-npu |
| 竞赛任务 | Track 1 模型适配 |
| 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。 |
| 项目 | 数值 |
|---|---|
| 模型仓库 | https://gitcode.com/nanyizjm/phikon-v2-npu |
| 原始模型或权重来源 | https://gitcode.com/hf_mirrors/owkin/phikon-v2 |
| 竞赛赛道 | Track 1:模型适配 |
| 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 |
| 精度要求 | NPU 结果与 CPU/GPU 参考值对比,误差小于 1% |
| 所需标签 | #+NPU |
| 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 |
| 资产或截图证明 | 已提供 |
README 必须包含明确的 CPU/GPU 与 NPU 数值对比数据。关键验收目标为误差小于 1%。相应的结构化证明在可用时应保存至 results/accuracy_eval.json 和 logs/accuracy_eval.log。
#+NPU
本部分直接写入 README 供平台审核使用。其仅使用本仓库中已签入的日志和 JSON 结果文件,不依赖嵌入式图片。
| 审核项 | 直接结果 |
|---|---|
| 仓库 | phikon-v2-npu |
| 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 - 选定的可复现误差 0.0005836880882270634% 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 | logs/inference.log、results/accuracy_eval.json、results/performance_eval.json、logs/accuracy_eval.log、logs/performance_eval.log |
2026-05-15 02:13:58,627 [INFO] Device: npu
2026-05-15 02:13:58,627 [INFO] Pooler output shape: (1, 1024)
2026-05-15 02:13:58,628 [INFO] Throughput: 73.04 img/s
"throughput_img_s": 73.0371735473345| 项目 | 数值 |
|---|---|
| 证据 | 在已检入的文本文件中未检测到 |
| 来源 | 指标 | 数值 |
|---|---|---|
results/accuracy_eval.json | pooler_output.relative_error.max_relative_error | 0.31620267033576965 |
results/accuracy_eval.json | pooler_output.relative_error.mean_relative_error | 0.003800167003646493 |
results/accuracy_eval.json | pooler_output.relative_error.filtered_mean_relative_error | 0.0009606852545402944 |
results/accuracy_eval.json | pooler_output.relative_error.max_absolute_error | 0.008023738861083984 |
results/accuracy_eval.json | pooler_output.relative_error.mean_absolute_error | 0.0005125721218064427 |
results/accuracy_eval.json | pooler_output.cosine_similarity | 0.9999974485804494 |
results/accuracy_eval.json | pooler_output.mre_pass | true |
results/accuracy_eval.json | pooler_output.cosine_pass | true |
results/accuracy_eval.json | last_hidden_state.relative_error.max_relative_error | 1 |
results/accuracy_eval.json | last_hidden_state.relative_error.mean_relative_error | 0.005364319775253534 |
精度结论:PASS - 选定的可复现误差 0.0005836880882270634% 低于 1%。
| 来源 | 指标 | 数值 |
|---|---|---|
results/performance_eval.json | device | npu |
results/performance_eval.json | dtype | float32 |
results/performance_eval.json | warmup | 3 |
results/performance_eval.json | num_runs | 10 |
results/performance_eval.json | results[0].batch_size | 1 |
results/performance_eval.json | results[0].mean_time_s | 0.012460637191543356 |
results/performance_eval.json | results[0].std_time_s | 0.00008950969657432262 |
results/performance_eval.json | results[0].p50_s | 0.012495438495534472 |
results/performance_eval.json | results[0].p95_s | 0.012541765489731915 |
results/performance_eval.json | results[0].throughput_img_s | 80.25271778867526 |
本文档记录 Phikon-v2 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。
Phikon-v2 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。
相关获取地址:
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。
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。
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。
| 组件 | 版本 / 说明 |
|---|---|
| 操作系统 | Linux 5.10.0-182.0.0.95.r2220_156.hce2.aarch64 aarch64 |
| Python | 3.11.14 |
| NPU 型号 | Ascend910_9362 |
| NPU 数量 | 2 |
| CANN | /usr/local/Ascend/cann-8.5.1 |
| PyTorch | 2.9.0+cpu |
| torch_npu | 2.9.0.post1+gitee7ba04 |
| transformers | 4.57.6 |
| 依赖安装 | pip install -r requirements.txt |
results/env_info.json 或 logs/env_check.log 为准)torch_npu,请先完成昇腾基础环境配置后再运行真实验证。.
├── .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/__init__.py
├── eval/eval_accuracy.py
├── eval/eval_performance.py
├── inference.py
├── locked_models.md
├── 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/performance_eval.json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pip install -r requirements.txt
python inference.py --model_path <model_path> --image_path <image.jpg> --device npupython eval/eval_accuracy.py --model_path <model_path> --device npu
python eval/eval_performance.py --model_path <model_path> --device npu| 指标 | 结果 |
|---|---|
| 模型名称 | phikon-v2 (Owkin phikon v2) |
| 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 | 1437334 |
测试结果来源:results/performance_eval.json
| 指标 | 结果 |
|---|---|
device | npu |
dtype | float32 |
image_size | 224 |
num_runs | 10 |
warmup | 3 |
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。
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。
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 |
--image_size | 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 |
python inference.py --help
python inference.py --model_path <model_path> --image_path <image.jpg> --device npu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。
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 |
|---|---|
| accuracy_eval_result | assets/accuracy_eval_result.png |
| env_check | assets/env_check.png |
| git_submit_result | assets/git_submit_result.png |
| inference_result | assets/inference_result.png |
| performance_eval_result | assets/performance_eval_result.png |
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。
# Environment Check Log
# Repository: phikon-v2-npu
# Model: phikon-v2 (Owkin phikon v2)
# Date: 2026-05-16 07:03:22
## System Info
Linux pod-8e032c81b34d489191e775768926f3b6 5.10.0-182.0.0.95.r2220_156.hce2.aarch64 #1 SMP Sat Sep 14 02:34:54 UTC 2024 aarch64 aarch64 aarch64 GNU/Linux
## Python
Python 3.11.14
pip 26.0.1 from /usr/local/python3.11.14/lib/python3.11/site-packages/pip (python 3.11)
## 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) |
+===========================+===============+====================================================+
| 0 Ascend910 | OK | 175.6 48 0 / 0 |
| 0 0 | 0000:0A:00.0 | 0 0 / 0 3107 / 65536 |
+------------------------------------------------------------------------------------------------+
| 0 Ascend910 | OK | - 49 0 / 0 |
| 1 1 | 0000:0B:00.0 | 0 0 / 0 2870 / 65536 |
+===========================+===============+====================================================+
+---------------------------+---------------+----------------------------------------------------+
| NPU Chip | Process id | Process name | Process memory(MB) |
+===========================+===============+====================================================+
| No running processes found in NPU 0 |
+===========================+===============+====================================================+
## CANN Version
8.5.1
## PyTorch
2.9.0+cpu
## torch_npu
2.9.0.post1+gitee7ba04
## transformers
4.57.6
## Git Info
Branch: main
Commit: 00a2f116639823f56b648538568f38067019a895
<redacted sensitive line>
ASCEND_TOOLKIT_HOME=/usr/local/Ascend/cann-8.5.1
PYTHONPATH=/usr/local/Ascend/cann-8.5.1/python/site-packages:/usr/local/Ascend/cann-8.5.1/opp/built-in/op_impl/ai_core/tbe:/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:/usr/local/Ascend/ascend-toolkit/latest/opp/built-in/op_impl/ai_core/tbe:{
"os": "Linux 5.10.0-182.0.0.95.r2220_156.hce2.aarch64 aarch64",
"python_version": "3.11.14",
"architecture": "aarch64",
"npu_model": "Ascend910_9362",
"npu_count": 2,
"cann_version": "/usr/local/Ascend/cann-8.5.1",
"torch_version": "2.9.0+cpu",
"torch_npu_version": "2.9.0.post1+gitee7ba04",
"transformers_version": "4.57.6",
"numpy_version": "1.26.4"
}2026-05-15 02:13:53,539 [INFO] Loading phikon-v2 from ./weights ...
2026-05-15 02:13:58,382 [INFO] Model loaded on npu, dtype=torch.float32
2026-05-15 02:13:58,384 [INFO] Parameters: 303,351,808 (303.4M)
2026-05-15 02:13:58,384 [INFO] Loading image: test_image_phikon.png
2026-05-15 02:13:58,390 [INFO] Input shape: torch.Size([1, 3, 224, 224])
2026-05-15 02:13:58,390 [INFO] Warming up ...
2026-05-15 02:13:58,577 [INFO] Running inference ...
2026-05-15 02:13:58,627 [INFO] === phikon-v2 Inference Results ===
2026-05-15 02:13:58,627 [INFO] Device: npu
2026-05-15 02:13:58,627 [INFO] Dtype: float32
2026-05-15 02:13:58,627 [INFO] Image size: 224x224
2026-05-15 02:13:58,627 [INFO] Pooler output shape: (1, 1024)
2026-05-15 02:13:58,627 [INFO] Last hidden state shape: (1, 197, 1024)
2026-05-15 02:13:58,627 [INFO] Pooler L2 norm: 9.6298
2026-05-15 02:13:58,628 [INFO] Hidden states mean L2 norm: 10.4621
2026-05-15 02:13:58,628 [INFO] Inference time: 0.0137s
2026-05-15 02:13:58,628 [INFO] Throughput: 73.04 img/s
2026-05-15 02:13:58,628 [INFO] Results: {
"device": "npu",
"dtype": "float32",
"image_size": 224,
"image_path": "test_image_phikon.png",
"pooler_shape": [
1,
1024
],
"last_hidden_shape": [
1,
197,
1024
],
"pooler_norm": 9.629796028137207,
"hidden_mean_norm": 10.462120056152344,
"inference_time_s": 0.013691657979507,
"throughput_img_s": 73.0371735473345
}2026-05-15 02:14:43,118 [INFO] Loading phikon-v2 model ...
2026-05-15 02:14:43,252 [INFO] Loading image: test_image_phikon.png
2026-05-15 02:14:43,257 [INFO] Running CPU inference ...
2026-05-15 02:14:45,579 [INFO] CPU pooler: (1, 1024), hidden: (1, 197, 1024)
2026-05-15 02:14:45,579 [INFO] Running NPU inference ...
2026-05-15 02:14:47,506 [INFO] NPU pooler: (1, 1024), hidden: (1, 197, 1024)
2026-05-15 02:14:47,506 [INFO] === Accuracy Comparison (NPU vs CPU) ===
2026-05-15 02:14:47,506 [INFO] Pooler Filtered MRE: 0.0961%
2026-05-15 02:14:47,506 [INFO] Pooler Cosine Similarity: 0.999997
2026-05-15 02:14:47,509 [INFO] Hidden States Filtered MRE: 0.0584%
2026-05-15 02:14:47,509 [INFO] Hidden States Cosine Similarity: 0.999998
2026-05-15 02:14:47,509 [INFO] Pooler MRE < 1%: PASS
2026-05-15 02:14:47,509 [INFO] Pooler Cosine > 0.9999: PASS
2026-05-15 02:14:47,509 [INFO] Hidden MRE < 1%: PASS
2026-05-15 02:14:47,509 [INFO] Hidden Cosine > 0.9999: PASS
2026-05-15 02:14:47,509 [INFO] Overall: PASS
2026-05-15 02:14:47,509 [INFO] Results saved to results/accuracy_eval.json{
"pooler_output": {
"relative_error": {
"max_relative_error": 0.31620267033576965,
"mean_relative_error": 0.003800167003646493,
"filtered_mean_relative_error": 0.0009606852545402944,
"max_absolute_error": 0.008023738861083984,
"mean_absolute_error": 0.0005125721218064427
},
"cosine_similarity": 0.9999974485804494,
"mre_pass": true,
"cosine_pass": true
},
"last_hidden_state": {
"relative_error": {
"max_relative_error": 1.0,
"mean_relative_error": 0.005364319775253534,
"filtered_mean_relative_error": 0.0005836880882270634,
"max_absolute_error": 0.019988059997558594,
"mean_absolute_error": 0.0004697777912952006
},
"cosine_similarity": 0.999998280389233,
"mre_pass": true,
"cosine_pass": true
},
"overall": "PASS"
}2026-05-15 02:15:17,103 [INFO] Loading phikon-v2 from ./weights ...
2026-05-15 02:15:19,104 [INFO] Model loaded on npu, dtype=torch.float32
2026-05-15 02:15:19,105 [INFO] Image size: 224, Batch sizes: [1, 2, 4, 8]
2026-05-15 02:15:19,105 [INFO] Warmup: 3, Num runs: 10
2026-05-15 02:15:19,105 [INFO]
--- Batch size: 1 ---
2026-05-15 02:15:19,515 [INFO] Mean: 0.0125s, P50: 0.0125s, P95: 0.0125s
2026-05-15 02:15:19,516 [INFO] Throughput: 80.25 img/s, NPU Memory: 1159.0 MB
2026-05-15 02:15:19,516 [INFO]
--- Batch size: 2 ---
2026-05-15 02:15:19,682 [INFO] Mean: 0.0124s, P50: 0.0124s, P95: 0.0125s
2026-05-15 02:15:19,683 [INFO] Throughput: 161.25 img/s, NPU Memory: 1160.4 MB
2026-05-15 02:15:19,683 [INFO]
--- Batch size: 4 ---
2026-05-15 02:15:19,880 [INFO] Mean: 0.0149s, P50: 0.0149s, P95: 0.0150s
2026-05-15 02:15:19,880 [INFO] Throughput: 267.83 img/s, NPU Memory: 1160.7 MB
2026-05-15 02:15:19,880 [INFO]
--- Batch size: 8 ---
2026-05-15 02:15:20,202 [INFO] Mean: 0.0245s, P50: 0.0244s, P95: 0.0246s
2026-05-15 02:15:20,202 [INFO] Throughput: 326.88 img/s, NPU Memory: 1163.0 MB
2026-05-15 02:15:20,202 [INFO]
=== Performance Summary ===
2026-05-15 02:15:20,202 [INFO] Batch Mean(s) P50(s) P95(s) img/s Mem(MB)
2026-05-15 02:15:20,202 [INFO] 1 0.0125 0.0125 0.0125 80.25 1159.0
2026-05-15 02:15:20,202 [INFO] 2 0.0124 0.0124 0.0125 161.25 1160.4
2026-05-15 02:15:20,202 [INFO] 4 0.0149 0.0149 0.0150 267.83 1160.7
2026-05-15 02:15:20,202 [INFO] 8 0.0245 0.0244 0.0246 326.88 1163.0
2026-05-15 02:15:20,203 [INFO] Results saved to results/performance_eval.json{
"device": "npu",
"dtype": "float32",
"image_size": 224,
"warmup": 3,
"num_runs": 10,
"results": [
{
"batch_size": 1,
"image_size": 224,
"mean_time_s": 0.012460637191543356,
"std_time_s": 8.950969657432262e-05,
"p50_s": 0.012495438495534472,
"p95_s": 0.012541765489731915,
"throughput_img_s": 80.25271778867526,
"npu_memory_mb": 1158.98193359375
},
{
"batch_size": 2,
"image_size": 224,
"mean_time_s": 0.01240277500182856,
"std_time_s": 5.324208228129922e-05,
"p50_s": 0.012393834491376765,
"p95_s": 0.012483979093667585,
"throughput_img_s": 161.25423541950386,
"npu_memory_mb": 1160.40673828125
},
{
"batch_size": 4,
"image_size": 224,
"mean_time_s": 0.014934969900059514,
"std_time_s": 1.9925904980612428e-05,
"p50_s": 0.014930603487300687,
"p95_s": 0.014963142519991379,
"throughput_img_s": 267.8277912019133,
"npu_memory_mb": 1160.70458984375
},
{
"batch_size": 8,
"image_size": 224,
"mean_time_s": 0.02447408049774822,
"std_time_s": 6.24867170506765e-05,
"p50_s": 0.024437255997327156,
"p95_s": 0.02456661449105013,
"throughput_img_s": 326.8764275224172,
"npu_memory_mb": 1163.00146484375
}
]
}license 元数据或 LICENSE 文件为准。