vllm-project/vllm

[Feature]: Default eplb num_redundant_experts to the lowest valid value if unspecified

Open

#30075 opened on Dec 4, 2025

View on GitHub
 (9 comments) (0 reactions) (0 assignees)Python (80,034 stars) (16,816 forks)batch import
feature requestgood first issuehelp wantedunstale

Description

🚀 The feature, motivation and pitch

EPLB requires the number of experts to be chosen up front and there is a known minimum valid value that can be derived from the vllm startup configuration. Since extra EPLB experts trades kv cache memory for potential performance improvements, but that is not guaranteed to pay off, having the EPLB value default to the minimum valid value would reduce friction on enabling EPLB the first time until users are ready to tune.

As a consequence, it would also streamline templating the same config to work across multiple EP sizes for the default case.

Alternatives

No response

Additional context

No response

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.

Contributor guide

Tech stack
python
Domain
machine learningaiperformance
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
half day
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
basic Pythonunderstanding of LLM serving config
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
40
Research direction
Start by examining the EPLB configuration initialization code in vllm/config.py or similar files. Identify where `num redundant experts` is currently set and derive the minimum valid value from the model configuration and expert parallelism settings. Then implement a fallback to this minimum when the parameter is not explicitly provided. Check existing tests for EPLB to ensure the new default does not break any.