What's the prefix for "Helsinki-NLP/opus-mt-zh-en" when using "AutoModelForSeq2SeqLM"?

[blackli7]

I was trying to translate a large amount of text in Chinese into English. So I tried to use "AutoModelForSeq2SeqLM" instead of "Pipeline" to run parallelism on gpus. But I found that this was a different between the two results.

When using "Pipeline" as follows:
pipe1 = pipeline("translation", model="Helsinki-NLP/opus-mt-zh-en") outputs1 = pipe1("比特币系统性能的428倍！世界第三代公链拥有方与上海这所高校达成合作") print(outputs1)
I got "That's 428 times the amount of Bitcoin's systematic power! The third generation of public chain owners in the world has made a partnership with Shanghai's high school."

However when using "AutoModelForSeq2SeqLM":
tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-zh-en") model = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-zh-en") encoded_input = tokenizer(["比特币系统性能的428倍！世界第三代公链拥有方与上海这所高校达成合作"], return_tensors="pt", padding=True, truncation=True, max_length=512) output = model.generate(**encoded_input, max_new_tokens=128, num_beams=1, num_return_sequences=1) output = tokenizer.batch_decode(output, skip_special_tokens=True) print(output)
I got "The world's third generation public chain owners have worked with Shanghai's college."

It seems "AutoModelForSeq2SeqLM" just translate the later part of the Chinese sentence. I guess it is because the prefix for this task is missing, but I find "model.config.prefix" is None. Could you show me the prefix for this task and this model, or could you help me do translation in a right way with "AutoModelForSeq2SeqLM"?

The colab link is https://colab.research.google.com/drive/1OTi_Nc8x1UTFUufbx7HsJfhn4VCUCkAs?usp=sharing