[InternVL 2.0] config of eof is INVALID, output token length is max_tokens

[spacegrass]

we set pad_id and end_id, but it does NOT work.
If we set max_tokens == 2048, then output token length is 2048
If we set max_tokens == 1024, then output token length is 1024
Inference time is very long ...
Even our question is: Is there a car in the image ?

---- How could we delete the padding-words by config ?

[nv-guomingz]

@kaiyux would u plz take a look on this question?