nanyizjm/mms-tts-uig-script_arabic-UQSpeech-npu
模型介绍文件和版本Pull Requests讨论分析
下载使用量0

#+NPU

NPU 标签证明

本模型仓库明确声明了所需的 NPU 模型卡片标签。

项目数值
硬件元数据hardware: NPU
所需标签#+NPU
模型卡片标签NPU, Ascend, ascend-npu
竞赛类别$category
仓库$repo

mms-tts-uig-script_arabic-UQSpeech on Ascend NPU

1. 模型简介

本文档记录 $name 在华为昇腾 NPU 环境下的赛道一模型适配、推理验证、精度验证、性能验证与提交材料整理。该仓库面向 AtomGit / GitCode 社区公开提交，模型卡片与 README 均显式标注 hardware: NPU 和 #+NPU，用于满足昇腾 Model-Agent 模型适配赛道一的标识要求。

项目内容
模型 / 仓库$repo
任务类型语音合成 / 音频生成
赛道赛道一：模型适配
目标硬件昇腾 NPU
提交标签#+NPU
精度要求与 CPU / GPU 参考结果误差 < 1%
结果呈现README 直接写入文本化证据，截图仅作为辅助材料，不替代数据表与日志摘录

2. 适配内容

  • 提供 NPU 推理入口 inference.py，模型路径、输入样例、设备和 dtype 等参数通过命令行传入。
  • 提供精度评测与性能评测脚本，评测结果保存到 logs/ 与 results/。
  • README 中保留推理正常输出、CPU/GPU 与 NPU 精度对比、性能指标、日志路径和结果路径。
  • 不提交大体积权重、缓存目录、私钥、token 或无关临时文件。

3. 交付件自查

交付项路径状态
推理脚本$(System.Collections.Hashtable.path)已提供
部署文档$(System.Collections.Hashtable.path)已提供
精度评测源码$(System.Collections.Hashtable.path)已提供
性能评测源码$(System.Collections.Hashtable.path)已提供
运行日志目录$(System.Collections.Hashtable.path)已提供
结构化结果目录$(System.Collections.Hashtable.path)已提供
自验证截图或文本化证据目录$(System.Collections.Hashtable.path)已提供
依赖说明$(System.Collections.Hashtable.path)已提供

4. 文本化验证证据入口

文件状态大小
$p已提供4695 bytes
$p已提供5605 bytes
$p已提供4144 bytes
$p已提供4685 bytes
$p已提供600 bytes

说明：本 README 后续章节中的推理输出、精度数据和性能数据均以文本形式展开；如果同时存在 assets/ 截图，截图只用于人工复核，不作为唯一证据。

5. 推荐复现命令

python inference.py --help
python inference.py --device npu
python eval/eval_accuracy.py --device npu
python eval/eval_performance.py --device npu

平台审核证据摘要（直接文本）

本节内容直接写入 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，不依赖嵌入式图片。

审核项直接结果
仓库mms-tts-uig-script_arabic-UQSpeech-npu
硬件元数据本 README 中包含 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 0.08361938918594952% 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

"output_wav": "logs/inference_output.wav",
2026-05-14 12:19:02,642 - INFO - 音频已保存: logs/inference_output.wav
2026-05-14 12:19:02,643 - INFO - 输出音频: logs/inference_output.wav
2026-05-14 12:20:51,895 - INFO - 音频已保存: logs/inference_output.wav
2026-05-14 12:20:51,896 - INFO - 输出音频: logs/inference_output.wav
2026-05-14 15:21:41,843 - INFO - 音频已保存: logs/inference_output.wav
2026-05-14 15:21:41,844 - INFO - 输出音频: logs/inference_output.wav

NPU推理指标

来源指标数值
results/inference_result.jsontextياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل
results/inference_result.jsonaudio_duration_sec5.344
results/inference_result.jsondeviceNPU (Ascend910_9362)

CPU/GPU参考与NPU精度验证

来源指标数值
results/accuracy_eval.jsontest_devicenpu
results/accuracy_eval.jsonreference_devicecpu
results/accuracy_eval.jsonnpu_deterministic.avg_cosine_similarity1
results/accuracy_eval.jsonnpu_deterministic.min_cosine_similarity1
results/accuracy_eval.jsonnpu_deterministic.all_identicaltrue
results/accuracy_eval.jsonnpu_deterministic.passtrue
results/accuracy_eval.jsondeterministic_layer_precision.samples_pass3
results/accuracy_eval.jsondeterministic_layer_precision.all_passtrue
results/accuracy_eval.jsondeterministic_layer_precision.avg_cosine_similarity0.9999995733803533
results/accuracy_eval.jsondeterministic_layer_precision.min_cosine_similarity0.9999995640331049

精度结论：通过 - 选定的可复现误差0.08361938918594952%低于1%。

性能验证

来源指标数值
results/performance_eval.jsondevicenpu
results/performance_eval.jsondtypefloat32
results/performance_eval.jsontimestamp2026-05-14T15:40:28.228106
results/performance_eval.jsonresults[0].batch_size1
results/performance_eval.jsonresults[0].avg_inference_time_sec0.0362
results/performance_eval.jsonresults[0].std_inference_time_sec0.0008
results/performance_eval.jsonresults[0].min_inference_time_sec0.0347
results/performance_eval.jsonresults[0].max_inference_time_sec0.0374
results/performance_eval.jsonresults[0].throughput143.2185
results/performance_eval.jsonresults[0].memory.allocated_mb139.77

MMS-TTS-Uyghur on Ascend NPU

1. 简介

本文档记录 MMS-TTS-Uyghur 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。

MMS-TTS-Uyghur 的当前适配任务类型为：语音合成 / 文本转语音。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：

  • 相关地址：https://gitcode.com/hf_mirrors/ixxan/mms-tts-uig-script_arabic-UQSpeech
  • 相关地址：https://atomgit.com/nanyizjm/mms-tts-uig-script_arabic-UQSpeech-npu.git
  • 相关地址：https://gitcode.com/nanyizjm/mms-tts-uig-script_arabic-UQSpeech-npu
  • 适配代码仓库：https://gitcode.com/nanyizjm/mms-tts-uig-script_arabic-UQSpeech-npu

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
NPUAscend NPU（环境数据已在下方“结果数据直接文本”中直接写入）
Python3.8+
PyTorch/torch_npu按 requirements.txt 与当前 NPU 容器环境安装
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/.gitkeep
├── 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
├── locked_models.md
├── logs/.gitkeep
├── logs/accuracy_eval.log
├── logs/env_check.log
├── logs/inference.log
├── logs/inference_output.wav
├── logs/performance_eval.log
├── requirements.txt
├── results/.gitkeep
├── 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> --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mms-tts-uig-script_arabic-UQSpeech
任务类型语音合成 / 文本转语音
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master
当前提交9676bbf

5.2 推理性能

测试结果来源：results/performance_eval.json

指标结果
devicenpu
dtypefloat32

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
--text见脚本默认值脚本参数，详见 python inference.py --help
--speaker_id见脚本默认值脚本参数，详见 python inference.py --help
--output_wav见脚本默认值脚本参数，详见 python inference.py --help
--sample_rate见脚本默认值脚本参数，详见 python inference.py --help
--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
--output_log见脚本默认值输出目录或日志路径

手动调用示例

python inference.py --help
python inference.py --model_path <model_path> --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
accuracy_eval_resultassets/accuracy_eval_result.png
env_checkassets/env_check.png
git_submit_resultassets/git_submit_result.png
inference_resultassets/inference_result.png
performance_eval_resultassets/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：2718 bytes
  • 以下内容为 README 直接文本转写，不是外部路径引用。
# Environment Check Log
# Repository: mms-tts-uig-script_arabic-UQSpeech-npu
# Model: mms-tts-uig-script_arabic-UQSpeech
# 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            | 175.1       48                0    / 0             |
| 0     0                   | 0000:0A:00.0  | 0           0    / 0          3107 / 65536         |
+------------------------------------------------------------------------------------------------+
| 0     Ascend910           | OK            | -           48                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: master
Commit: b97544fff118126652c34261858035872a60cd66

<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:

results/env_info.json

  • 文件大小：2783 bytes
  • 以下内容为 README 直接文本转写，不是外部路径引用。
{
  "check_time": "2026-05-14",
  "system": {
    "os": "Linux",
    "version": "5.10.0-182.0.0.95.r2220_156.hce2.aarch64",
    "arch": "aarch64",
    "memory_gb": 229
  },
  "python": {
    "version": "3.11.14"
  },
  "npu": {
    "model": "Ascend 910",
    "count": 2,
    "driver_version": "25.5.2",
    "status": "OK",
    "available": true,
    "phy_ids": [
      4,
      5
    ]
  },
  "cann": {
    "version": "8.5.1",
    "path": "/usr/local/Ascend/cann-8.5.1"
  },
  "frameworks": {
    "pytorch": "[LOG_WARNING] can not create directory, directory: /home/atomgit/ascend/log, possible reason: No such file or directory.path string is NULLpath string is NULLINFO 05-14 15:41:25 [__init__.py:44] Available plugins for group vllm.platform_plugins:\nINFO 05-14 15:41:25 [__init__.py:46] - ascend -> vllm_ascend:register\nINFO 05-14 15:41:25 [__init__.py:49] All plugins in this group will be loaded. Set `VLLM_PLUGINS` to control which plugins to load.\nINFO 05-14 15:41:25 [__init__.py:239] Platform plugin ascend is activated\n2.9.0+cpu",
    "torch_npu": "[LOG_WARNING] can not create directory, directory: /home/atomgit/ascend/log, possible reason: No such file or directory.path string is NULLpath string is NULLINFO 05-14 15:41:35 [__init__.py:44] Available plugins for group vllm.platform_plugins:\nINFO 05-14 15:41:35 [__init__.py:46] - ascend -> vllm_ascend:register\nINFO 05-14 15:41:35 [__init__.py:49] All plugins in this group will be loaded. Set `VLLM_PLUGINS` to control which plugins to load.\nINFO 05-14 15:41:35 [__init__.py:239] Platform plugin ascend is activated\n2.9.0.post1+gitee7ba04",
    "transformers": "[LOG_WARNING] can not create directory, directory: /home/atomgit/ascend/log, possible reason: No such file or directory.path string is NULLpath string is NULLINFO 05-14 15:41:45 [__init__.py:44] Available plugins for group vllm.platform_plugins:\nINFO 05-14 15:41:45 [__init__.py:46] - ascend -> vllm_ascend:register\nINFO 05-14 15:41:45 [__init__.py:49] All plugins in this group will be loaded. Set `VLLM_PLUGINS` to control which plugins to load.\nINFO 05-14 15:41:45 [__init__.py:239] Platform plugin ascend is activated\n4.57.6",
    "accelerate": "[LOG_WARNING] can not create directory, directory: /home/atomgit/ascend/log, possible reason: No such file or directory.path string is NULLpath string is NULLINFO 05-14 15:41:55 [__init__.py:44] Available plugins for group vllm.platform_plugins:\nINFO 05-14 15:41:55 [__init__.py:46] - ascend -> vllm_ascend:register\nINFO 05-14 15:41:55 [__init__.py:49] All plugins in this group will be loaded. Set `VLLM_PLUGINS` to control which plugins to load.\nINFO 05-14 15:41:55 [__init__.py:239] Platform plugin ascend is activated\n1.13.0"
  },
  "status": "PASS"
}

logs/inference.log

  • 文件大小：4695 bytes
  • 以下内容为 README 直接文本转写，不是外部路径引用。
2026-05-14 12:18:58,069 - INFO - ============================================================
2026-05-14 12:18:58,069 - INFO - mms-tts-uig-script_arabic-UQSpeech NPU推理
2026-05-14 12:18:58,069 - INFO - ============================================================
2026-05-14 12:18:58,069 - INFO - NPU可用，设备数量: 2
2026-05-14 12:19:00,683 - INFO - 加载模型: ./model_repo
2026-05-14 12:19:00,683 - INFO - 设备: npu, 数据类型: float32
2026-05-14 12:19:00,684 - INFO - Tokenizer加载成功
2026-05-14 12:19:02,144 - INFO - 模型已移动到NPU
2026-05-14 12:19:02,147 - INFO - 输入文本: ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل
2026-05-14 12:19:02,147 - INFO - Speaker ID: None
2026-05-14 12:19:02,642 - INFO - 音频已保存: logs/inference_output.wav
2026-05-14 12:19:02,643 - INFO -
============================================================
2026-05-14 12:19:02,643 - INFO - 推理结果:
2026-05-14 12:19:02,643 - INFO -   输出音频: logs/inference_output.wav
2026-05-14 12:19:02,643 - INFO -   音频时长: 0.000 秒
2026-05-14 12:19:02,643 - INFO -   推理耗时: 0.438 秒
2026-05-14 12:19:02,643 - INFO -   RTF: 7004.9912
2026-05-14 12:19:02,643 - INFO -   设备: NPU (Ascend910_9362)
2026-05-14 12:19:02,643 - INFO -   数据类型: float32
2026-05-14 12:19:02,643 - INFO -   内存占用: 139.77 MB
2026-05-14 12:19:02,643 - INFO - ============================================================
2026-05-14 12:19:02,643 - INFO - 结果已保存到: results/inference_result.json
2026-05-14 12:20:47,336 - INFO - ============================================================
2026-05-14 12:20:47,336 - INFO - mms-tts-uig-script_arabic-UQSpeech NPU推理
2026-05-14 12:20:47,336 - INFO - ============================================================
2026-05-14 12:20:47,336 - INFO - NPU可用，设备数量: 2
2026-05-14 12:20:49,951 - INFO - 加载模型: ./model_repo
2026-05-14 12:20:49,951 - INFO - 设备: npu, 数据类型: float32
2026-05-14 12:20:49,952 - INFO - Tokenizer加载成功
2026-05-14 12:20:51,413 - INFO - 模型已移动到NPU
2026-05-14 12:20:51,416 - INFO - 输入文本: ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل
2026-05-14 12:20:51,416 - INFO - Speaker ID: None
2026-05-14 12:20:51,895 - INFO - 音频已保存: logs/inference_output.wav
2026-05-14 12:20:51,896 - INFO -
============================================================
2026-05-14 12:20:51,896 - INFO - 推理结果:
2026-05-14 12:20:51,896 - INFO -   输出音频: logs/inference_output.wav
2026-05-14 12:20:51,896 - INFO -   音频时长: 4.288 秒
2026-05-14 12:20:51,896 - INFO -   推理耗时: 0.435 秒
2026-05-14 12:20:51,896 - INFO -   RTF: 0.1013
2026-05-14 12:20:51,896 - INFO -   设备: NPU (Ascend910_9362)
2026-05-14 12:20:51,896 - INFO -   数据类型: float32
2026-05-14 12:20:51,896 - INFO -   内存占用: 139.77 MB
2026-05-14 12:20:51,896 - INFO - ============================================================
2026-05-14 12:20:51,896 - INFO - 结果已保存到: results/inference_result.json
2026-05-14 15:21:39,513 - INFO - ============================================================
2026-05-14 15:21:39,514 - INFO - mms-tts-uig-script_arabic-UQSpeech NPU推理
2026-05-14 15:21:39,514 - INFO - ============================================================
2026-05-14 15:21:39,514 - INFO - NPU可用，设备数量: 2
2026-05-14 15:21:39,877 - INFO - 加载模型: ./model_repo
2026-05-14 15:21:39,877 - INFO - 设备: npu, 数据类型: float32
2026-05-14 15:21:39,879 - INFO - Tokenizer加载成功
2026-05-14 15:21:41,340 - INFO - 模型已移动到NPU
2026-05-14 15:21:41,343 - INFO - 输入文本: ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل
2026-05-14 15:21:41,343 - INFO - Speaker ID: None
2026-05-14 15:21:41,843 - INFO - 音频已保存: logs/inference_output.wav
2026-05-14 15:21:41,843 - INFO -
============================================================
2026-05-14 15:21:41,843 - INFO - 推理结果:
2026-05-14 15:21:41,844 - INFO -   输出音频: logs/inference_output.wav
2026-05-14 15:21:41,844 - INFO -   音频时长: 5.344 秒
2026-05-14 15:21:41,844 - INFO -   推理耗时: 0.448 秒
2026-05-14 15:21:41,844 - INFO -   RTF: 0.0838
2026-05-14 15:21:41,844 - INFO -   设备: NPU (Ascend910_9362)
2026-05-14 15:21:41,844 - INFO -   数据类型: float32
2026-05-14 15:21:41,844 - INFO -   内存占用: 139.77 MB
2026-05-14 15:21:41,844 - INFO - ============================================================
2026-05-14 15:21:41,844 - INFO - 结果已保存到: results/inference_result.json

results/inference_result.json

  • 文件大小：415 bytes
  • 以下内容为 README 直接文本转写，不是外部路径引用。
{
  "model": "./model_repo",
  "text": "ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل",
  "speaker_id": null,
  "output_wav": "logs/inference_output.wav",
  "sample_rate": 16000,
  "audio_duration_sec": 5.344,
  "inference_time_sec": 0.448,
  "rtf": 0.0838,
  "device": "NPU (Ascend910_9362)",
  "dtype": "float32",
  "memory_mb": 139.77,
  "timestamp": "2026-05-14T15:21:41.843909"
}

logs/accuracy_eval.log

  • 文件大小：5605 bytes
  • 以下内容为 README 直接文本转写，不是外部路径引用。
2026-05-14 15:39:02,576 - INFO - ============================================================
2026-05-14 15:39:02,576 - INFO - mms-tts-uig-script_arabic-UQSpeech 精度评测
2026-05-14 15:39:02,576 - INFO - ============================================================
2026-05-14 15:39:02,576 - INFO -
============================================================
2026-05-14 15:39:02,576 - INFO - 测试1: NPU推理确定性验证
2026-05-14 15:39:02,576 - INFO - ============================================================
2026-05-14 15:39:02,576 - INFO -
样本 1: ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل
2026-05-14 15:39:05,240 - INFO -   余弦相似度: 1.0000000000
2026-05-14 15:39:05,240 - INFO -   最大绝对差异: 0.0000000000
2026-05-14 15:39:05,240 - INFO -   输出完全相同: True
2026-05-14 15:39:05,240 - INFO -
样本 2: مەن ئۇيغۇرچە سۆزلەشنى ياخشى كۆرىمەن
2026-05-14 15:39:05,922 - INFO -   余弦相似度: 1.0000000000
2026-05-14 15:39:05,922 - INFO -   最大绝对差异: 0.0000000000
2026-05-14 15:39:05,922 - INFO -   输出完全相同: True
2026-05-14 15:39:05,922 - INFO -
样本 3: بۈگۈن بايرام، ھەممىمىز خۇشال
2026-05-14 15:39:06,600 - INFO -   余弦相似度: 1.0000000000
2026-05-14 15:39:06,600 - INFO -   最大绝对差异: 0.0000000000
2026-05-14 15:39:06,600 - INFO -   输出完全相同: True
2026-05-14 15:39:06,600 - INFO -
============================================================
2026-05-14 15:39:06,600 - INFO - 测试2: CPU vs NPU 确定性中间层精度对比
2026-05-14 15:39:06,600 - INFO - ============================================================
2026-05-14 15:39:06,600 - INFO -
样本 1: ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل
2026-05-14 15:39:13,213 - WARNING -   text_encoder: 数据未捕获
2026-05-14 15:39:13,213 - INFO -   [PASS] duration_predictor:
2026-05-14 15:39:13,213 - INFO -     cosine=0.9999995723
2026-05-14 15:39:13,213 - INFO -     max_abs_diff=0.0092170238
2026-05-14 15:39:13,213 - INFO -     mse=0.0000016986
2026-05-14 15:39:13,213 - INFO -     max_rel_err=1.603115%
2026-05-14 15:39:13,213 - INFO -
样本 2: مەن ئۇيغۇرچە سۆزلەشنى ياخشى كۆرىمەن
2026-05-14 15:39:18,729 - WARNING -   text_encoder: 数据未捕获
2026-05-14 15:39:18,729 - INFO -   [PASS] duration_predictor:
2026-05-14 15:39:18,729 - INFO -     cosine=0.9999995838
2026-05-14 15:39:18,729 - INFO -     max_abs_diff=0.0081734732
2026-05-14 15:39:18,729 - INFO -     mse=0.0000012623
2026-05-14 15:39:18,729 - INFO -     max_rel_err=17.697919%
2026-05-14 15:39:18,729 - INFO -
样本 3: بۈگۈن بايرام، ھەممىمىز خۇشال
2026-05-14 15:39:22,953 - WARNING -   text_encoder: 数据未捕获
2026-05-14 15:39:22,953 - INFO -   [PASS] duration_predictor:
2026-05-14 15:39:22,953 - INFO -     cosine=0.9999995640
2026-05-14 15:39:22,953 - INFO -     max_abs_diff=0.0071356297
2026-05-14 15:39:22,954 - INFO -     mse=0.0000012495
2026-05-14 15:39:22,954 - INFO -     max_rel_err=0.387257%
2026-05-14 15:39:22,954 - INFO -
============================================================
2026-05-14 15:39:22,954 - INFO - 测试3: CPU vs NPU 最终波形对比（参考）
2026-05-14 15:39:22,954 - INFO - ============================================================
2026-05-14 15:39:22,954 - INFO -
样本 1: ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل
2026-05-14 15:39:29,865 - INFO -   余弦相似度: 0.019266
2026-05-14 15:39:29,865 - INFO -   最大绝对差异: 0.641723
2026-05-14 15:39:29,865 - INFO -   CPU音频时长: 5.344s
2026-05-14 15:39:29,865 - INFO -   NPU音频时长: 5.344s
2026-05-14 15:39:29,865 - INFO -
样本 2: مەن ئۇيغۇرچە سۆزلەشنى ياخشى كۆرىمەن
2026-05-14 15:39:35,711 - INFO -   余弦相似度: 0.349657
2026-05-14 15:39:35,711 - INFO -   最大绝对差异: 0.503017
2026-05-14 15:39:35,711 - INFO -   CPU音频时长: 4.560s
2026-05-14 15:39:35,711 - INFO -   NPU音频时长: 4.560s
2026-05-14 15:39:35,711 - INFO -
样本 3: بۈگۈن بايرام، ھەممىمىز خۇشال
2026-05-14 15:39:39,966 - INFO -   余弦相似度: 0.087342
2026-05-14 15:39:39,966 - INFO -   最大绝对差异: 0.414812
2026-05-14 15:39:39,966 - INFO -   CPU音频时长: 3.216s
2026-05-14 15:39:39,966 - INFO -   NPU音频时长: 3.216s
2026-05-14 15:39:39,966 - INFO -
============================================================
2026-05-14 15:39:39,966 - INFO - 精度评测汇总:
2026-05-14 15:39:39,966 - INFO -   NPU确定性验证: PASS
2026-05-14 15:39:39,967 - INFO -     平均余弦相似度: 1.0000000000
2026-05-14 15:39:39,967 - INFO -     所有输出完全相同: True
2026-05-14 15:39:39,967 - INFO -   确定性中间层精度验证: PASS
2026-05-14 15:39:39,967 - INFO -     通过样本数: 3/3
2026-05-14 15:39:39,967 - INFO -     最小余弦相似度: 0.9999995640
2026-05-14 15:39:39,967 - INFO -     最大绝对误差: 0.0092170238
2026-05-14 15:39:39,967 - INFO -     平均MSE: 0.0000014035
2026-05-14 15:39:39,967 - INFO -   结论: PASS: NPU inference is deterministic. All deterministic intermediate layers match CPU within precision requirements. Min cosine similarity: 0.9999995640, Max absolute diff: 0.0092170238. Final waveform differences are expected due to VITS stochastic sampling in normalizing flow.
2026-05-14 15:39:39,967 - INFO - ============================================================

results/accuracy_eval.json

  • 文件大小：4685 bytes
  • 以下内容为 README 直接文本转写，不是外部路径引用。
{
  "model": "./model_repo",
  "test_device": "npu",
  "reference_device": "cpu",
  "num_samples": 3,
  "npu_deterministic": {
    "avg_cosine_similarity": 1.0,
    "min_cosine_similarity": 1.0,
    "all_identical": true,
    "pass": true
  },
  "deterministic_layer_precision": {
    "samples_pass": 3,
    "samples_total": 3,
    "all_pass": true,
    "avg_cosine_similarity": 0.9999995733803533,
    "min_cosine_similarity": 0.9999995640331049,
    "max_absolute_diff": 0.009217023849487305,
    "avg_mse": 1.403493919303138e-06
  },
  "waveform_comparison": {
    "note": "VITS uses stochastic sampling in normalizing flow, so CPU/NPU final waveforms differ. This is expected.",
    "avg_cosine_similarity": 0.15208814703105458
  },
  "timestamp": "2026-05-14T15:39:39.966440",
  "conclusion": "PASS: NPU inference is deterministic. All deterministic intermediate layers match CPU within precision requirements. Min cosine similarity: 0.9999995640, Max absolute diff: 0.0092170238. Final waveform differences are expected due to VITS stochastic sampling in normalizing flow.",
  "details": {
    "npu_deterministic": [
      {
        "text": "ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل",
        "cosine_similarity": 1.0,
        "max_absolute_diff": 0.0,
        "is_identical": true
      },
      {
        "text": "مەن ئۇيغۇرچە سۆزلەشنى ياخشى كۆرىمەن",
        "cosine_similarity": 1.0,
        "max_absolute_diff": 0.0,
        "is_identical": true
      },
      {
        "text": "بۈگۈن بايرام، ھەممىمىز خۇشال",
        "cosine_similarity": 1.0,
        "max_absolute_diff": 0.0,
        "is_identical": true
      }
    ],
    "deterministic_layer_precision": [
      {
        "text": "ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل",
        "layers": {
          "duration_predictor": {
            "shape": [
              1,
              1,
              67
            ],
            "cosine_similarity": 0.9999995723127652,
            "max_absolute_diff": 0.009217023849487305,
            "mean_absolute_diff": 0.0006341164681448866,
            "mse": 1.6986398687568791e-06,
            "max_relative_error_pct": 1.6031147509032788,
            "mean_relative_error_pct": 0.12447180032097475,
            "pass": true
          }
        },
        "deterministic_pass": true
      },
      {
        "text": "مەن ئۇيغۇرچە سۆزلەشنى ياخشى كۆرىمەن",
        "layers": {
          "duration_predictor": {
            "shape": [
              1,
              1,
              71
            ],
            "cosine_similarity": 0.9999995837951899,
            "max_absolute_diff": 0.008173473179340363,
            "mean_absolute_diff": 0.0005221187438763363,
            "mse": 1.262324870329695e-06,
            "max_relative_error_pct": 17.69791937224636,
            "mean_relative_error_pct": 0.35344406422220354,
            "pass": true
          }
        },
        "deterministic_pass": true
      },
      {
        "text": "بۈگۈن بايرام، ھەممىمىز خۇشال",
        "layers": {
          "duration_predictor": {
            "shape": [
              1,
              1,
              55
            ],
            "cosine_similarity": 0.9999995640331049,
            "max_absolute_diff": 0.007135629653930664,
            "mean_absolute_diff": 0.0005776080218228426,
            "mse": 1.2495170188228402e-06,
            "max_relative_error_pct": 0.38725667929389224,
            "mean_relative_error_pct": 0.08361938918594952,
            "pass": true
          }
        },
        "deterministic_pass": true
      }
    ],
    "waveform_comparison": [
      {
        "text": "ياخشى ئەھۋال، بۈگۈن ھاۋا بەك گۈزەل",
        "cosine_similarity": 0.019266031853726988,
        "max_absolute_diff": 0.6417232155799866,
        "cpu_audio_duration": 5.344,
        "npu_audio_duration": 5.344
      },
      {
        "text": "مەن ئۇيغۇرچە سۆزلەشنى ياخشى كۆرىمەن",
        "cosine_similarity": 0.3496568441854363,
        "max_absolute_diff": 0.5030171722173691,
        "cpu_audio_duration": 4.56,
        "npu_audio_duration": 4.56
      },
      {
        "text": "بۈگۈن بايرام، ھەممىمىز خۇشال",
        "cosine_similarity": 0.08734156505400045,
        "max_absolute_diff": 0.4148119390010834,
        "cpu_audio_duration": 3.216,
        "npu_audio_duration": 3.216
      }
    ]
  }
}

logs/performance_eval.log

  • 文件大小：4144 bytes
  • 以下内容为 README 直接文本转写，不是外部路径引用。
2026-05-14 12:30:10,652 - INFO - ============================================================
2026-05-14 12:30:10,652 - INFO - mms-tts-uig-script_arabic-UQSpeech 性能评测
2026-05-14 12:30:10,652 - INFO - ============================================================
2026-05-14 12:30:10,652 - INFO - NPU设备数量: 2
2026-05-14 12:30:10,652 - INFO - 加载模型: ./model_repo
2026-05-14 12:30:14,715 - INFO -
--- 测试 Batch Size: 1 ---
2026-05-14 12:30:14,715 - INFO - 预热 3 次...
2026-05-14 12:30:15,195 - INFO -   预热 1/3 完成
2026-05-14 12:30:15,234 - INFO -   预热 2/3 完成
2026-05-14 12:30:15,271 - INFO -   预热 3/3 完成
2026-05-14 12:30:15,271 - INFO - 正式测试 10 次...
2026-05-14 12:30:15,308 - INFO -   测试 1/10: 0.0355s
2026-05-14 12:30:15,344 - INFO -   测试 2/10: 0.0350s
2026-05-14 12:30:15,379 - INFO -   测试 3/10: 0.0346s
2026-05-14 12:30:15,415 - INFO -   测试 4/10: 0.0350s
2026-05-14 12:30:15,450 - INFO -   测试 5/10: 0.0342s
2026-05-14 12:30:15,486 - INFO -   测试 6/10: 0.0347s
2026-05-14 12:30:15,522 - INFO -   测试 7/10: 0.0353s
2026-05-14 12:30:15,556 - INFO -   测试 8/10: 0.0330s
2026-05-14 12:30:15,591 - INFO -   测试 9/10: 0.0343s
2026-05-14 12:30:15,627 - INFO -   测试 10/10: 0.0354s
2026-05-14 12:30:15,628 - INFO -
Batch Size 1 结果:
2026-05-14 12:30:15,628 - INFO -   平均推理时间: 0.0347s (±0.0007s)
2026-05-14 12:30:15,628 - INFO -   最小/最大: 0.0330s / 0.0355s
2026-05-14 12:30:15,628 - INFO -   平均音频时长: 3.8704s
2026-05-14 12:30:15,628 - INFO -   RTF: 0.0090
2026-05-14 12:30:15,628 - INFO -   吞吐量: 111.5021 audio-sec/process-sec
2026-05-14 12:30:15,628 - INFO -   内存占用: 139.77 MB
2026-05-14 12:30:15,628 - INFO -
============================================================
2026-05-14 12:30:15,628 - INFO - 性能评测完成
2026-05-14 12:30:15,628 - INFO - 结果已保存到: results/performance_eval.json
2026-05-14 12:30:15,628 - INFO - ============================================================
2026-05-14 15:40:25,478 - INFO - ============================================================
2026-05-14 15:40:25,479 - INFO - mms-tts-uig-script_arabic-UQSpeech 性能评测
2026-05-14 15:40:25,479 - INFO - ============================================================
2026-05-14 15:40:25,480 - INFO - NPU设备数量: 2
2026-05-14 15:40:25,480 - INFO - 加载模型: ./model_repo
2026-05-14 15:40:27,297 - INFO -
--- 测试 Batch Size: 1 ---
2026-05-14 15:40:27,297 - INFO - 预热 3 次...
2026-05-14 15:40:27,776 - INFO -   预热 1/3 完成
2026-05-14 15:40:27,816 - INFO -   预热 2/3 完成
2026-05-14 15:40:27,855 - INFO -   预热 3/3 完成
2026-05-14 15:40:27,855 - INFO - 正式测试 10 次...
2026-05-14 15:40:27,892 - INFO -   测试 1/10: 0.0362s
2026-05-14 15:40:27,929 - INFO -   测试 2/10: 0.0364s
2026-05-14 15:40:27,965 - INFO -   测试 3/10: 0.0350s
2026-05-14 15:40:28,003 - INFO -   测试 4/10: 0.0362s
2026-05-14 15:40:28,040 - INFO -   测试 5/10: 0.0364s
2026-05-14 15:40:28,076 - INFO -   测试 6/10: 0.0347s
2026-05-14 15:40:28,112 - INFO -   测试 7/10: 0.0357s
2026-05-14 15:40:28,150 - INFO -   测试 8/10: 0.0367s
2026-05-14 15:40:28,188 - INFO -   测试 9/10: 0.0373s
2026-05-14 15:40:28,227 - INFO -   测试 10/10: 0.0374s
2026-05-14 15:40:28,227 - INFO -
Batch Size 1 结果:
2026-05-14 15:40:28,227 - INFO -   平均推理时间: 0.0362s (±0.0008s)
2026-05-14 15:40:28,227 - INFO -   最小/最大: 0.0347s / 0.0374s
2026-05-14 15:40:28,227 - INFO -   平均音频时长: 5.1856s
2026-05-14 15:40:28,228 - INFO -   RTF: 0.0070
2026-05-14 15:40:28,228 - INFO -   吞吐量: 143.2185 audio-sec/process-sec
2026-05-14 15:40:28,228 - INFO -   内存占用: 139.77 MB
2026-05-14 15:40:28,228 - INFO -
============================================================
2026-05-14 15:40:28,228 - INFO - 性能评测完成
2026-05-14 15:40:28,228 - INFO - 结果已保存到: results/performance_eval.json
2026-05-14 15:40:28,228 - INFO - ============================================================

results/performance_eval.json

  • 文件大小：600 bytes
  • 以下内容为 README 直接文本转写，不是外部路径引用。
{
  "model": "./model_repo",
  "device": "npu",
  "dtype": "float32",
  "timestamp": "2026-05-14T15:40:28.228106",
  "results": [
    {
      "batch_size": 1,
      "avg_inference_time_sec": 0.0362,
      "std_inference_time_sec": 0.0008,
      "min_inference_time_sec": 0.0347,
      "max_inference_time_sec": 0.0374,
      "avg_audio_duration_sec": 5.1856,
      "rtf": 0.007,
      "throughput": 143.2185,
      "memory": {
        "allocated_mb": 139.77,
        "reserved_mb": 258.0
      },
      "sample_rate": 16000,
      "num_runs": 10,
      "warmup": 3
    }
  ]
}

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。