opus-mt-ine-en Ascend NPU 部署指南

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

docker exec -it test-modelagent bash

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

pip install transformers torch_npu sacremoses

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

python3 inference.py

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

python3 inference.py --precision_test

============================================================
opus-mt-ine-en Ascend NPU 部署测试
============================================================
MODEL_DIR: /data/ysws/agentsp/5-20-1/Helsinki-NLP/opus-mt-ine-en
OUTPUT_DIR: /data/ysws/agentsp/5-20-1/opus-mt-ine-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-ine-en-ascend/test_sentences.txt
共 8 句

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

测试句子数量: 8
  [1] Hallo, kako si hodie?
  [2] Jasne, dusma, adoro te.
  [3] Weather is beautiful hours.
  [4] 'Bunjķn, amigo.
  [5] Merci za pomoc.
  [6] Jabber ID de GnuPG
  [7] læro traduction machine.
  [8] Vi sei morre

开始翻译 (device: npu:0)...

翻译结果:
  [1] 原文: Hallo, kako si hodie?
      译文: Hello, how are you?
  [2] 原文: Jasne, dusma, adoro te.
      译文: Yeah, honey, I love you.
  [3] 原文: Weather is beautiful hours.
      译文: Weather is beautiful hours.
  [4] 原文: 'Bunjķn, amigo.
      译文: 'Bunion, amigo.
  [5] 原文: Merci za pomoc.
      译文: Thank you for your help.
  [6] 原文: Jabber ID de GnuPG
      译文: Jabber ID of GnuPG
  [7] 原文: læro traduction machine.
      译文: learn translation machine.
  [8] 原文: Vi sei morre
      译文: You're dying

总耗时: 1.2028s
平均每句: 0.1504s

============================================================
opus-mt-ine-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: Hallo, kako si hodie?
CPU 译文: Hello, how are you?
CPU 耗时: 1.5833s
NPU 译文: Hello, how are you?
NPU 耗时: 0.1461s
译文匹配: True

--- 句子 2 ---
原文: Jasne, dusma, adoro te.
CPU 译文: Yeah, honey, I love you.
CPU 耗时: 1.8263s
NPU 译文: Yeah, honey, I love you.
NPU 耗时: 0.1438s
译文匹配: True

--- 句子 3 ---
原文: Vi sei morre
CPU 译文: You're dying
CPU 耗时: 1.0393s
NPU 译文: You're dying
NPU 耗时: 0.0902s
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: 1.4829s
平均 NPU 推理时间: 0.1267s
NPU 加速比: 11.70x

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-ine-en"

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

texts = ["Hallo, kako si hodie?"]
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"],
        attention_mask=inputs["attention_mask"],
        max_length=100,
        num_beams=4,
        early_stopping=True
    )

translations = tokenizer.batch_decode(gen_ids, skip_special_tokens=True)
print(translations)  # ['Hello, how are you?']