opus-mt-en-ROMANCE Ascend NPU 部署指南

opus-mt-en-ROMANCE-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-en-ROMANCE-ascend/

# 普通推理 (仅测NPU)
python3 inference.py

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

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

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

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

测试句子数量: 8
  [1] I love you, but I love him more!
  [2] Hello, how are you today?
  [3] The quick brown fox jumps over the lazy dog.
  [4] This is a sample sentence for machine translation testing.
  [5] Good morning! Nice to meet you.
  [6] Thank you very much for your help.
  [7] What is the weather like today?
  [8] I am learning machine translation.

开始翻译 (device: npu:0)...

翻译结果:
  [1] 原文: I love you, but I love him more!
      译文: Te amo, pero lo amo más!
  [2] 原文: Hello, how are you today?
      译文: Bonjour, comment estás hoy?
  [3] 原文: The quick brown fox jumps over the lazy dog.
      译文: La raposa marrón rápida salta sobre el perro preguiçoso.
  [4] 原文: This is a sample sentence for machine translation testing.
      译文: Esta é unha frase de exemplo para tests de traducción automática.
  [5] 原文: Good morning! Nice to meet you.
      译文: - Encantado de conocerla.
  [6] 原文: Thank you very much for your help.
      译文: Muchas gracias por su ayuda.
  [7] 原文: What is the weather like today?
      译文: Como es el clima de hoy?
  [8] 原文: I am learning machine translation.
      译文: Estou aprendendo a traducción automatique.

总耗时: 1.2013s
平均每句: 0.1502s

============================================================
opus-mt-en-ROMANCE 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: I love you, but I love him more!
CPU 译文: Te amo, pero lo amo más!
CPU 耗时: 2.2079s
NPU 译文: Te amo, pero lo amo más!
NPU 耗时: 0.1707s
译文匹配: True

--- 句子 2 ---
原文: Hello, how are you today?
CPU 译文: Bonjour, comment estás hoy?
CPU 耗时: 1.5499s
NPU 译文: Bonjour, comment estás hoy?
NPU 耗时: 0.1120s
译文匹配: True

--- 句子 3 ---
原文: The quick brown fox jumps over the lazy dog.
CPU 译文: La raposa marrón rápida salta sobre el perro preguiçoso.
CPU 耗时: 3.6907s
NPU 译文: La raposa marrón rápida salta sobre el perro preguiçoso.
NPU 耗时: 0.2639s
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: 2.4828s
平均 NPU 推理时间: 0.1822s
NPU 加速比: 13.63x

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-en-ROMANCE"

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

texts = ["I love you, but I love him more!"]
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)  # ['Te amo, pero lo amo más!']

texts = [
    "I love you, but I love him more!",
    "Hello, how are you today?",
    "Thank you very much for your help."
]

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)
for src, trans in zip(texts, translations):
    print(f"{src} -> {trans}")

{
  "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,
  "max_position_embeddings": 512,
  "vocab_size": 65001
}