Hi all,

I am just exported model to ONNX format, via following script:

CONVERSION TO ONNX ------------------

import torch
from transformers import MBartForConditionalGeneration, MBart50TokenizerFast

model_path = "./mbart_large_50_model"
model = MBartForConditionalGeneration.from_pretrained(model_path)
tokenizer = MBart50TokenizerFast.from_pretrained(model_path)

input_text = "This is just simple text"
inputs = tokenizer(input_text, return_tensors="pt")

onnx_path = "./mbart_large_50.onnx"

• export to ONNX format
  torch.onnx.export(
  model, # PyTorch model
  (inputs["input_ids"],), # input data
  onnx_path, # Pth to the ONNX file
  input_names=["input_ids"], # input tensor names
  output_names=["logits"], # otput tensor names
  dynamic_axes={"input_ids": {0: "batch", 1: "sequence"}},
  opset_version=14
  )
  -----------------END OF CONVERSION-----------------

• got something, but I am not sure if I missed something?
  -Now, while dealing with original PyTorch model (so, not onnx exporter one), there is a way to specify source and target language for translation:
  ....
  model = MBartForConditionalGeneration.from_pretrained(model_path)
  tokenizer = MBart50TokenizerFast.from_pretrained(model_path)

• specify src language
  tokenizer.src_lang = "en_XX"
  ...
  generated_tokens = model.generate(**inputs, forced_bos_token_id=tokenizer.lang_code_to_id["hr_HR"])

-Pay attention on:
forced_bos_token_id=tokenizer.lang_code_to_id["hr_HR"]
-Wondering, is there a way (and HOW?) to do the same with exported ONNX model?

                                      #TRIED SO FAR---

tokenizer = MBart50TokenizerFast.from_pretrained(tokenizer_path)
tokenizer.src_lang = "en_XX"

input_text = "Hello, my name is BART"
inputs = tokenizer(input_text, return_tensors="np")

target_lang_id = tokenizer.lang_code_to_id["hr_HR"]
input_ids = np.array(inputs["input_ids"], dtype=np.int64)
start_token = np.array([[target_lang_id]], dtype=np.int64)
input_ids = np.concatenate((start_token, input_ids), axis=1)

-As you can see, I am trying to put the target_lang_id as first input entry.
but, this way only the first word is translated, not the rest.

-Any help?