Conformer repeats the same translation

[sarapapi]

What is your question?

I have trained the newly added Conformer model on MuST-C En-Es and I noticed that sometimes it starts to repeat the same words or part of sentence, here is an example: Primero, documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta primero. Primero, documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta primero. Primero documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta documenta And I discovered that this happens when the inference is done in batch: if I set --max-sentences to 1, everything works. Have you ever experienced this issue? Do you have any idea on how to solve it? Thanks

What's your environment?

• fairseq Version (e.g., 1.0 or main): main
• PyTorch Version (e.g., 1.0): 1.9
• OS (e.g., Linux): Linux
• How you installed fairseq (pip, source): source
• Python version: 3.8.13
• CUDA/cuDNN version: 11.6
• GPU models and configuration: tried on A40 and V100
• Any other relevant information: //

[fengpeng-yue]

I also had the same problem @sravyapopuri388 , can you give some advice please?

[85631486]

I also had the same problem, can you give some advice please?

[85631486]

I have trained a Englist to Chinese Model. I also had the same problem.

input: the platform truck can be used for transporting transport airplanes , such as Y-8 transport airplane , II-76MD transport airplane , etc

output: 可用于运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运运

This is unusual and sometimes recurring.

[85631486]

Most have no problems, and it's strange that they happen in individual cases.

Tensor2Tensor is no problems.

[85631486]

When input like below the result is normal. input: (the platform truck can be used for transporting transport airplanes , such as Y-8 transport airplane , II-76MD transport airplane , etc) output:（台车可用于运输运输机，如y-8运输机、ii-76md运输机等）

Why???