NPU Tag Evidence

This repository is published as an Ascend NPU model repository. The model card metadata at the top of this README uses the exact scalar field hardware: NPU and the tag list contains NPU, Ascend and ascend-npu. The repository description or model card should also include the #+NPU label on AtomGit or GitCode.

Item	Value
Repository	https://gitcode.com/nanyizjm/Qwen3-ForcedAligner-0.6B
Competition task	Track 1 model adaptation
Hardware metadata	hardware: NPU
Required tag	#+NPU
README data policy	Inference, accuracy and performance values are written as text in this README; images are not used as a replacement for data.

Track 1 Model Card Summary

Item	Value
Model repository	https://gitcode.com/nanyizjm/Qwen3-ForcedAligner-0.6B
Original model or weight source	https://gitcode.com/hf_mirrors/Qwen/Qwen3-ForcedAligner-0.6B
Competition track	Track 1: model adaptation
Target hardware	Ascend NPU
Required function	NPU inference runs successfully or the blocking reason is explicitly recorded
Required accuracy	NPU result compared with CPU/GPU reference, error less than 1 percent
Required tag	#+NPU

Deliverable Checklist

Deliverable	Status
inference.py	Present
readme.md / README.md	Present
eval/eval_accuracy.py	Present
eval/eval_performance.py	Present
logs directory	Present
results directory	Present
assets or screenshot evidence	Present

Accuracy Evidence Requirement

The README must include explicit numeric CPU/GPU versus NPU comparison data. The key acceptance target is error less than 1 percent. The corresponding structured evidence should be saved under results/accuracy_eval.json and logs/accuracy_eval.log when available.

#+NPU

Qwen3-ForcedAligner-0.6B on Ascend NPU

1. 简介

本文档记录 Qwen3-ForcedAligner-0.6B 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。

Qwen3-ForcedAligner-0.6B 的当前适配任务类型为：语音识别 / 音频理解。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
NPU 数量	2
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
├── 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/model_check.log
├── logs/performance_eval.log
├── requirements.txt
├── results/accuracy_eval.json
├── results/env_info.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 --audio <audio.wav> --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	`Qwen3-ForcedAligner-0.6B`
任务类型	语音识别 / 音频理解
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	`9f3093a`

5.2 推理性能

测试结果来源：results/performance_eval.json

指标	结果
`device`	npu:0
`num_runs`	5
`warmup`	2
`avg_latency_ms`	166.00

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
`--audio`	见脚本默认值	输入样例路径
`--asr-model`	见脚本默认值	脚本参数，详见 python inference.py --help
`--aligner-model`	见脚本默认值	脚本参数，详见 python inference.py --help
`--language`	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
`--analyze-only`	见脚本默认值	脚本参数，详见 python inference.py --help
`--output`	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 --audio <audio.wav> --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。

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

=== Environment Check ===
OS: Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35
Python: 3.11.14
Architecture: aarch64
Hostname: pod-8e032c81b34d489191e775768926f3b6
PyTorch: 2.9.0+cpu
torch_npu: 2.9.0.post1+gitee7ba04
transformers: 4.57.6
accelerate: 1.13.0
NPU available: True
NPU count: 2
NPU name: Ascend910_9362
CANN path: /usr/local/Ascend/cann-8.5.1
SOC version: ascend910_9391
NPU visible devices: 4,5

results/env_info.json

文件大小：504 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": "3.11.14",
  "arch": "aarch64",
  "hostname": "pod-8e032c81b34d489191e775768926f3b6",
  "torch": "2.9.0+cpu",
  "torch_npu": "2.9.0.post1+gitee7ba04",
  "npu_available": true,
  "npu_count": 2,
  "npu_name": "Ascend910_9362",
  "transformers": "4.57.6",
  "accelerate": "1.13.0",
  "cann_path": "/usr/local/Ascend/cann-8.5.1",
  "soc_version": "ascend910_9391",
  "npu_visible_devices": "4,5"
}

logs/model_check.log

文件大小：632 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

=== Model Weight Check ===
ForcedAligner path: /opt/atomgit/models/Qwen3-ForcedAligner-0.6B
ForcedAligner files: ['README.md', '._____temp', 'vocab.json', 'merges.txt', 'model.safetensors', 'config.json', 'preprocessor_config.json', '.msc', '.mv', 'generation_config.json', 'configuration.json', 'tokenizer_config.json', 'chat_template.json']
ASR path: /opt/atomgit/models/Qwen3-ASR-0.6B
ASR files: ['README.md', '._____temp', 'vocab.json', 'merges.txt', 'model.safetensors', 'config.json', 'preprocessor_config.json', '.msc', '.mv', 'generation_config.json', 'configuration.json', 'tokenizer_config.json', 'chat_template.json']

logs/inference.log

文件大小：1632 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 NULLINFO 05-14 14:21:13 [__init__.py:44] Available plugins for group vllm.platform_plugins:
INFO 05-14 14:21:13 [__init__.py:46] - ascend -> vllm_ascend:register
INFO 05-14 14:21:13 [__init__.py:49] All plugins in this group will be loaded. Set `VLLM_PLUGINS` to control which plugins to load.
INFO 05-14 14:21:13 [__init__.py:239] Platform plugin ascend is activated
WARNING 05-14 14:21:15 [registry.py:911] Model architecture Qwen3ASRForConditionalGeneration is already registered, and will be overwritten by the new model class <class 'qwen_asr.core.vllm_backend.qwen3_asr.Qwen3ASRForConditionalGeneration'>.
[INFO] NPU available: True
[INFO] NPU count: 2
[INFO] NPU name: Ascend910_9362
[INFO] Loading ASR model...
The following generation flags are not valid and may be ignored: ['temperature']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
[INFO] Model loaded in 5.2s
[INFO] Running inference on test audio...
Setting `pad_token_id` to `eos_token_id`:151645 for open-end generation.
/opt/atomgit/.local/lib/python3.11/site-packages/qwen_asr/core/transformers_backend/modeling_qwen3_asr.py:314: UserWarning: Cannot create tensor with interal format while allow_internel_format=False, tensor will be created with base format. (Triggered internally at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:338.)
  input_lengths_leave = input_lengths % 100
[INFO] Inference time: 1196ms
[INFO] Language:
[INFO] Text: ""
[INFO] NPU inference test PASSED

logs/accuracy_eval.log

文件大小：214 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

Accuracy Evaluation: NPU vs CPU
Audio: /opt/atomgit/models/test_audio.wav
CPU text: "" | NPU text: ""
Text match: True
Language match: True
Timestamp count: CPU=0 NPU=0
Error rate: 0.0%
Accuracy pass: True

results/accuracy_eval.json

文件大小：410 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{
  "cpu_text": "",
  "npu_text": "",
  "cpu_language": "",
  "npu_language": "",
  "text_match": true,
  "language_match": true,
  "cpu_timestamp_count": 0,
  "npu_timestamp_count": 0,
  "timestamp_count_match": true,
  "timestamp_avg_deviation_ms": 0,
  "timestamp_max_deviation_ms": 0,
  "cpu_latency_ms": 4712,
  "npu_latency_ms": 678,
  "accuracy_pass": true,
  "error_rate_percent": 0.0
}

logs/performance_eval.log

文件大小：156 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

Performance Evaluation
Device: npu:0
Runs: 5, Warmup: 2
Avg latency: 166.0ms
P50: 166.4ms, P90: 167.5ms
Throughput: 6.025 req/s
NPU memory: 1497.7MB

results/performance_eval.json

文件大小：395 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{
  "device": "npu:0",
  "num_runs": 5,
  "warmup": 2,
  "avg_latency_ms": 166.0,
  "min_latency_ms": 164.0,
  "max_latency_ms": 167.7,
  "p50_latency_ms": 166.4,
  "p90_latency_ms": 167.5,
  "p99_latency_ms": 167.7,
  "std_latency_ms": 1.5,
  "throughput_rps": 6.025,
  "npu_memory_mb": 1497.7,
  "latencies_ms": [
    167.2,
    167.7,
    166.4,
    164.5,
    164.0
  ]
}

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。