[Performance]: first token latency during inference is longer when the number of input tokens is small

[gaochenxi]

Proposal to improve performance

Describe the question
In vLLM, I noticed that the first token latency during inference is longer when the number of input tokens is small (e.g., 30 tokens), compared to when the number of input tokens is large (e.g., 800 tokens).

Is this expected behavior? Could you help explain why this happens?

Environment

vLLM version: (v0.8.4)

Model: (DeepSeek-R1)

GPU: (H200 * 8)

Report of performance regression

No response

Misc discussion on performance

No response

Your current environment (if you think it is necessary)

The output of `python collect_env.py`

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.