liucu/SampurNER_Bengali_MuRIL
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

prachuryyaIITG/SampurNER_Bengali_MuRIL on Ascend NPU

模型与任务

  • 模型:prachuryyaIITG/SampurNER_Bengali_MuRIL
  • Revision:65b89c6ae58f7bb910ca82365e0e1e71dc2cdc4f
  • 适配路线:text-sequence-labeling
  • 设备:npu:0

环境依赖

{
  "device": "Ascend910_9362",
  "torch": "2.9.0+cpu",
  "torch_npu": "2.9.0.post1+gitee7ba04",
  "transformers": "5.9.0",
  "python": "3.11.14",
  "os": "Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35",
  "dtype": "float32"
}

安装:pip install -r requirements.txt

NPU 推理

默认执行:python inference.py

真实输入

{
  "type": "real Bengali text sequences (fine-grained NER fixtures)",
  "sentences": 3,
  "padded_shape": [
    3,
    18
  ],
  "real_tokens": 44,
  "max_length": 128,
  "padding": "right, pad_token_id=0",
  "note": "模型实际为孟加拉语文本序列标注(MuRIL 底座 BERT token 分类),任务类别按真实契约执行"
}

真实推理结果

{
  "logits_shape": [
    3,
    18,
    133
  ],
  "labels": 133,
  "decoded_entities_sentence0": [
    {
      "type": "person-artist/author",
      "text": "রবীন্দ্রনাথঠাকুর"
    },
    {
      "type": "location-GPE",
      "text": "কলকাতায়"
    },
    {
      "type": "organization-education",
      "text": "শান্তিনিকেতন"
    }
  ]
}

CPU-NPU 一致性

{
  "compared_elements": 5852,
  "mask": "attention_mask real-token positions only",
  "max_abs_diff": 0.05438041687011719,
  "mean_abs_diff": 0.003917993501998307,
  "label_agreement": 1.0,
  "all_finite": true,
  "criterion": "max_abs_diff <= 0.05 * std(cpu_logits)=1.6760 -> bound=0.0838(基于 CPU 基线实测输出尺度制定)",
  "tool": "scripts/compare_outputs.py",
  "passed": true
}

性能

{
  "warmup_iters": 3,
  "timed_iters": 10,
  "per_iter_ms": 14.11,
  "sentences_per_s": 212.54,
  "tokens_per_s": 3117.3,
  "peak_device_memory_mb": 1946.0,
  "timing": "torch.npu.synchronize() before/after"
}

自验证截图

Agent workflow

NPU device call

Model result

三张图片均由 xterm.js 根据本次真实日志生成。

已知限制

  • 仅对 revision 65b89c6 的权重、fp32 dtype、3 条固定孟加拉语输入与 Ascend910 环境完成验证。
  • 一致性为 smoke consistency:CPU 与 NPU logits 对比 + 100% 标签一致率,非完整评测集 F1 复测。
  • 该模型为文本细粒度 NER(token 分类),非 DNA/RNA 生物序列架构;按真实契约 text-sequence-labeling 路线适配。

标签

#NPU #Ascend #Ascend910