How can I speed up using llmlingua2 ?

[yyjabiding]

Describe the issue

I have a context length of about 100k.
Is there any methods that I can do to speed up using llmlingua2 to compress it and keep it within s short time like shorter than 2 seconds?"
Thanks.

[iofu728]

Hi @yyjabiding, thanks for your interest in LLMLingua.

Although we haven't tested it, it seems possible. LLMLingua-2 forwards a BERT-level model chunk by chunk, so increasing the batch size could potentially reduce latency. You can check the implementation here: LLMLingua Prompt Compressor.