[Usage]: What should I do if I want to skip the prefill of a new request?

[chenhongyu2048]

Your current environment

The output of `python collect_env.py`

How would you like to use vllm

MyQuestion:
I want to add a new request to engine, and make some dummy kv cache for it, and then let it directly start the decode stage.
What should I do for this?
Maybe I should first modify the STATUS in SeqGroup, allocate and faked some kv cache in the block manager?
But i'm still confused about how to add the faked first output token for it?

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.