Motivation

There are quite a few near misses that cause the test to fail; relaxing some thresholds to reduce test failures from near misses. Models adjusted:

• zai-org/GLM-4.5V-FP8
• moonshotai/Kimi-VL-A3B-Instruct
• Qwen/Qwen2.5-VL-7B-Instruct
• OpenGVLab/InternVL2_5-2B
• unsloth/Mistral-Small-3.1-24B-Instruct-2503

re @merrymercy

reduce the threshold for all failed nightly-test-vlm-accuracy-2-gpu-runnerhttps://github.com/sgl-project/sglang/actions/runs/20513077000/job/58936524276#step:4:6851

Modifications

Increased latency threshold and lowered score threshold slightly.

Accuracy Tests

(https://github.com/sgl-project/sglang/actions/runs/20764464710/job/59627330523)

Benchmarking and Profiling

n/a

Checklist

• [Done ] Format your code according to the Format code with pre-commit.
• [ Done] Add unit tests according to the Run and add unit tests.
• [ Done] Update documentation according to Write documentations.
• [Done ] Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• [Done ] Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments (/tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci) or contact authorized users to do so.
4. After green CI and required approvals, ask Merge Oncalls to merge.