support for token-classification

### Feature request

how to support a token-classification model such as `microsoft/llmlingua-2-xlm-roberta-large-meetingbank`

### Motivation

I want to do prompt compression.

### Your contribution

I am pleased to support a PR if you could give me some hint.