opus-mt-en-zh 是 Helsinki-NLP 系列的多语言翻译模型,专门用于英语(EN)到中文(ZH)的翻译任务,采用 6 层 Transformer 编码器 + 6 层解码器架构,参数量约 74M。
opus-mt-en-zh-ascend/
├── inference.py # 推理测试脚本
├── log.txt # 测试日志
├── README.md # 本文档
├── test_sample.txt # 测试样例
├── inference_result.json # 推理结果
└── precision_result.json # 精度测试结果docker exec -it test-modelagent bashsource /usr/local/Ascend/ascend-toolkit/set_env.sh模型文件位于 /data/ysws/agentsp/5-17-1/opus-mt-en-zh/ 目录下:
pip install transformers torch_npu -i https://huaweimirror.com.cn/simple运行推理脚本进行翻译:
cd /data/ysws/agentsp/5-17-1/opus-mt-en-zh-ascend/
python3 inference.py运行精度对比测试,验证 NPU 计算结果与 CPU 一致性:
cd /data/ysws/agentsp/5-17-1/opus-mt-en-zh-ascend/
python3 inference.py precision_test| 指标 | 实测值 | 阈值 | 状态 |
|---|---|---|---|
| 翻译一致性 | 100% | 100% | PASS |
| 输出匹配 | True | True | PASS |
| 操作 | 耗时 |
|---|---|
| CPU 推理时间 | 2.425s |
| NPU 推理时间 | 0.163s |
| 加速比 | 14.83x |
| 输入 (英语) | 输出 (中文) |
|---|---|
| "Hello, how are you today?" | "你好,你今天好吗?" |
| "I am very happy to see you." | Chinese translation |
| "The weather is nice today." | Chinese translation |
结果: CPU 和 NPU 输出的翻译结果完全一致
============================================================
OPUS-MT-EN-ZH NPU Test
Model: Helsinki-NLP/opus-mt-en-zh (EN → ZH)
Output: /data/ysws/agentsp/5-17-1/opus-mt-en-zh-ascend
============================================================
============================================================
OPUS-MT-EN-ZH Inference Test (NPU)
============================================================
Device: npu:0
Model: /data/ysws/agentsp/5-17-1/opus-mt-en-zh/Helsinki-NLP/opus-mt-en-zh
Loading tokenizer...
Loading model...
Loading weights: 100%|██████████| 258/258 [00:00<00:00, 11224.72it/s]
Input text: ['Hello, how are you today?']
Input shape: torch.Size([1, 8])
Generated text: ['你好,你今天好吗?']
Inference time: 1.100s
============================================================
Precision Test (CPU vs NPU)
============================================================
Loading tokenizer...
Loading model on CPU...
Loading weights: 100%|██████████| 258/258 [00:00<00:00, 11175.69it/s]
Running inference on CPU...
Loading model on NPU...
Loading weights: 100%|██████████| 258/258 [00:00<00:00, 11194.99it/s]
Running inference on NPU...
CPU inference time: 2.425s
NPU inference time: 0.163s
Speedup: 14.83x
CPU output: ['你好,你今天好吗?']
NPU output: ['你好,你今天好吗?']
Output texts match: True
Status: PASS
============================================================
Test Complete!
============================================================import torch
from transformers import MarianMTModel, MarianTokenizer
MODEL_DIR = "/data/ysws/agentsp/5-17-1/opus-mt-en-zh/Helsinki-NLP/opus-mt-en-zh"
tokenizer = MarianTokenizer.from_pretrained(MODEL_DIR)
model = MarianMTModel.from_pretrained(MODEL_DIR)
model = model.to("npu:0")
model.eval()
src_texts = ["Hello, how are you today?"]
inputs = tokenizer(src_texts, return_tensors="pt", padding=True)
inputs = {k: v.to("npu:0") for k, v in inputs.items()}
with torch.no_grad():
outputs = model.generate(inputs['input_ids'], max_new_tokens=50)
translations = tokenizer.batch_decode(outputs, skip_special_tokens=True)
print(translations) # ['你好,你今天好吗?']src_texts = [
"Hello, how are you today?",
"I am very happy to see you.",
"The weather is nice today."
]
inputs = tokenizer(src_texts, return_tensors="pt", padding=True)
inputs = {k: v.to("npu:0") for k, v in inputs.items()}
with torch.no_grad():
outputs = model.generate(inputs['input_ids'], max_new_tokens=50)
translations = tokenizer.batch_decode(outputs, skip_special_tokens=True)
for src, tgt in zip(src_texts, translations):
print(f"{src} -> {tgt}")从 config.json 提取的关键参数:
{
"vocab_size": 58101,
"d_model": 512,
"encoder_layers": 6,
"decoder_layers": 6,
"encoder_attention_heads": 8,
"decoder_attention_heads": 8,
"encoder_ffn_dim": 2048,
"decoder_ffn_dim": 2048
}A: 检查 NPU 驱动是否正确安装,确保 CANN 环境变量已 source。
A: 使用批处理可以显著提高吞吐量。另外,首次推理会有编译开销,后续推理会更快。
A: 本模型专门用于英语到中文的翻译。如需其他语言对,请访问 Helsinki-NLP 模型库。
本项目遵循 Apache-2.0 许可证