OPUS-MT-UR-EN Ascend NPU 部署指南

opus-mt-ur-en-ascend/
├── inference.py          # 推理测试脚本
├── log.txt               # 测试日志
├── README.md             # 本文档
├── test_sample.txt       # 测试样本
├── inference_result.json # 推理结果
└── 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

cd /data/ysws/agentsp/5-18-1/opus-mt-ur-en-ascend/

# 使用默认测试句子
python3 inference.py

# 指定设备
python3 inference.py npu:0

cd /data/ysws/agentsp/5-18-1/opus-mt-ur-en-ascend/

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

============================================================
OPUS-MT-UR-EN NPU Test
============================================================
Device: npu:0
Model: /data/ysws/agentsp/5-18-1/opus-mt-ur-en/Helsinki-NLP/opus-mt-ur-en

Input text: ['ہیلو، آپ کیسے ہیں؟']
Input shape: torch.Size([1, 7])
Generated text: ['Hello, how are you?']
Inference time: 1.202s

============================================================
Precision Test (CPU vs NPU)
============================================================
CPU inference time: 1.729s
NPU inference time: 0.135s
Speedup: 12.81x
CPU output: ['Hello, how are you?']
NPU output: ['Hello, how are you?']
Output texts match: True
Status: PASS

============================================================
Test Sample
============================================================
  1. ہیلو، آپ کیسے ہیں؟
  2. آپ کو دیکھ کر بہت خوشی ہوئی
  3. آج موسم اچھا ہے
============================================================

import torch
from transformers import MarianTokenizer, MarianMTModel

MODEL_DIR = "/data/ysws/agentsp/5-18-1/opus-mt-ur-en/Helsinki-NLP/opus-mt-ur-en"

tokenizer = MarianTokenizer.from_pretrained(MODEL_DIR)
model = MarianMTModel.from_pretrained(MODEL_DIR)
model = model.to("npu:0")
model.eval()

src_texts = ["ہیلو، آپ کیسے ہیں؟"]
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)  # ["Hello, how are you?"]

src_texts = [
    "ہیلو، آپ کیسے ہیں؟",
    "مجھے آپ سے مل کر بہت خوشی ہوئی",
    "آج کا موسم بہت اچھا ہے"
]

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

{
  "hidden_size": 768,
  "encoder_layers": 6,
  "decoder_layers": 6,
  "encoder_attention_heads": 12,
  "decoder_attention_heads": 12,
  "d_model": 768
}