opus-mt-sem-en Ascend NPU 部署指南

opus-mt-sem-en-ascend/
├── inference.py          # 推理测试脚本
├── log.txt               # 测试日志
├── README.md             # 本文档
├── precision_result.json # 精度测试结果
└── test_sentences.txt   # 测试句子

docker exec -it test-modelagent bash

source /usr/local/Ascend/ascend-toolkit/set_env.sh

pip install transformers torch_npu

cd /data/ysws/agentsp/5-20-1/opus-mt-sem-en-ascend/

# 使用默认测试句子
python3 inference.py

cd /data/ysws/agentsp/5-20-1/opus-mt-sem-en-ascend/

# 运行完整精度测试
python3 inference.py --precision_test

============================================================
opus-mt-sem-en Ascend NPU 部署测试
============================================================
MODEL_DIR: /data/ysws/agentsp/5-20-1/Helsinki-NLP/opus-mt-sem-en
OUTPUT_DIR: /data/ysws/agentsp/5-20-1/opus-mt-sem-en-ascend
Mode: precision_test

============================================================
创建测试样本
============================================================
测试句子已保存到: /data/ysws/agentsp/5-20-1/opus-mt-sem-en-ascend/test_sentences.txt
共 8 句

============================================================
opus-mt-sem-en NPU 推理测试
============================================================
Device: npu:0
Model loaded successfully!

测试句子数量: 8
  [1] Dobry den, jak se mate?
  [2] Jake je pocasi?
  [3] Dekuji mockrat.
  [4] Na shledanou!
  [5] Kolik to stoji?
  [6] Nevim.
  [7] Jsem student.
  [8] Good morning!

开始翻译 (device: npu:0)...

翻译结果:
  [1] 原文: Dobry den, jak se mate?
      译文: Dobry den, would you mate?
  [2] 原文: Jake je pocasi?
      译文: Jacke jee pocasi?
  [3] 原文: Dekuji mockrat.
      译文: Mocks.
  [4] 原文: Na shledanou!
      译文: We're not schledanou!
  [5] 原文: Kolik to stoji?
      译文: Collick tostories?
  [6] 原文: Nevim.
      译文: Nevim.
  [7] 原文: Jsem student.
      译文: Name a student.
  [8] 原文: Good morning!
      译文: Good Morning!

总耗时: 1.5194s
平均每句: 0.1899s

============================================================
opus-mt-sem-en 精度测试 (CPU vs NPU)
============================================================
Device: npu:0

加载 CPU 模型...
CPU 模型加载完成

加载 NPU 模型...
NPU 模型加载完成

测试句子数量: 3

--- 句子 1 ---
原文: Dobry den, jak se mate?
CPU 译文: Dobry den, would you mate?
CPU 耗时: 3.1759s
NPU 译文: Dobry den, would you mate?
NPU 耗时: 0.2323s
译文匹配: True

--- 句子 2 ---
原文: Jake je pocasi?
CPU 译文: Jacke jee pocasi?
CPU 耗时: 2.1074s
NPU 译文: Jacke jee pocasi?
NPU 耗时: 0.1572s
译文匹配: True

--- 句子 3 ---
原文: Na shledanou!
CPU 译文: We're not schledanou!
CPU 耗时: 2.3885s
NPU 译文: We're not schledanou!
NPU 耗时: 0.1763s
译文匹配: True

============================================================
精度测试结果汇总
============================================================
译文完全匹配: PASS
平均 CPU 推理时间: 2.5573s
平均 NPU 推理时间: 0.1886s
NPU 加速比: 13.56x

精度阈值: 1.0%
译文匹配率: PASS

总体状态: PASS

============================================================
测试完成!
============================================================

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

MODEL_DIR = "/data/ysws/agentsp/5-20-1/Helsinki-NLP/opus-mt-sem-en"

tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_DIR)
model = model.to("npu:0")
model.eval()

texts = ["Dobry den, jak se mate?"]
inputs = tokenizer(texts, return_tensors="pt", padding=True)
inputs = {k: v.to("npu:0") for k, v in inputs.items()}

with torch.no_grad():
    gen_ids = model.generate(inputs["input_ids"], max_length=100, num_beams=4, early_stopping=True)
    translations = tokenizer.batch_decode(gen_ids, skip_special_tokens=True)

print(translations)  # ['Dobry den, would you mate?']