BartForConditionalGeneration with PubMedBERT/Bio_ClinicalBERT tokenizer

I am using BartForConditionalGeneration architecture from @sgugger to train the model to summarize medical reports. The model achieves good results when the tokenizer is the same as the model. However, I tried to replace the tokenizer with a PubMedBERT/Bio_ClinicalBERT tokenizer, but the model doesn’t learn well.
Do you have any idea why the model doesn’t work well with any of these medical tokenizers? Or what will be a good way to use PubMedBERT/Bio_ClinicalBERT tokenizer in the BartForConditionalGeneration model?

Thank you very much! I appreciate your help!