vllm-project/vllm

[Feature]: Run performance benchmarks for multi-modal models in CI

Open

#16,353 opened on Apr 9, 2025

View on GitHub
 (15 comments) (0 reactions) (0 assignees)Python (80,034 stars) (16,816 forks)batch import
feature requesthelp wantedkeep-openmulti-modality

Description

🚀 The feature, motivation and pitch

We currently only have benchmarks for text-only models such as Llama. With the increasing importance of multi-modality and related optimizations such as processor cache, we should add performance benchmarks for multi-modal models to avoid regressions (e.g. memory leaks, slow batching).

We can measure the peak memory usage based on this code:

import resource

max_self_usage = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / (1 << 20)
max_children_usage = resource.getrusage(resource.RUSAGE_CHILDREN).ru_maxrss / (1 << 20)
print(f"Peak memory usage: {max_self_usage} (self) + {max_children_usage} (children) GiB")

Alternatives

No response

Additional context

cc @mgoin @ywang96

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 learningperformance
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.
1-2 days
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
Understanding of multi modal modelsFamiliarity with vllm codebaseExperience with CI pipelinesKnowledge of Python benchmarking
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
50
Research direction
Examine the existing text only benchmarks in the repository (likely under tests/benchmarks). Study how multi modal models are integrated in vllm (e.g., processor cache handling). Implement performance benchmarks that measure peak memory using resource.getrusage as shown in the issue. Ensure the benchmarks run in CI (e.g., via pytest or custom workflow). Consult the linked PR or comments from maintainers @mgoin and @ywang96 for specific expectations.