Conversion innaccuracy specific Opus-MT model

[FricoRico]

System Info

python==3.12.4

onnxruntime==1.19
onnxruntime-tools==1.7

optimum[exporters]==1.22
transformers==4.44

Who can help?

@michaelbenayoun @JingyaHuang @echarlaix

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction (minimal, reproducible, runnable)

I'm converting several Opus-MT (translation) models to ONNX format, after which I validate the model by doing a translation. I've noticed weird behavior with Helsinki-NLP/opus-tatoeba-en-ja specifically. When running the model the output keeps repeating the same tokens several times. I suspect this might have something to do with the torch model being in float16 format, whilst all other models I have tried are in float32. See torch_dtype in config.json.

Input sentence:

The company revealed three deals, including an investment in startup X-Energy and two development agreements that aim to add around 300 megawatts worth of capacity in both the Pacific Northwest and Virginia, two data center hotspots.

Output sentence:
同社 3 件 スタートアップ スタートアップ スタートアップ 2015 同社 同社 同社 同社 同社 同社 3 件 3 件 3 件 株式 3 件 スタートアップ 3 件 スタートアップ 3 件 スタートアップ 3 件 スタートアップ 3 件 スタートアップ 3 件 スタートアップ 3 件 スタートアップ 3 件 3

Expected behavior

Just like converting to and from other languages I expect the input sentence to be correctly translated into an output sentence.