[Klaud Cold] Update kimik2.5-fp4-b300-vllm vLLM image to v0.21.0

[functionstackx]

Summary

• Bumps kimik2.5-fp4-b300-vllm from vllm/vllm-openai:v0.19.0-cu130 (27d old) to vllm/vllm-openai:v0.21.0.

Test plan

• Full sweep passes with full-sweep-enabled label.

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26006165007
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26006165007

[claude]

Straightforward vLLM image version bump for kimik2.5-fp4-b300-vllm with corresponding changelog entry.

Extended reasoning...

Overview

This PR updates a single config entry in .github/configs/nvidia-master.yaml to bump the vLLM image for kimik2.5-fp4-b300-vllm from v0.19.0-cu130 to v0.21.0, and adds a matching entry to perf-changelog.yaml. The scope is minimal — two lines of config change plus a 6-line changelog entry.

Security risks

None. This is a config-only change pointing to an upstream vLLM image tag from the same vllm/vllm-openai registry path already used by sibling recipes. No code, auth, crypto, or permission paths are touched.

Level of scrutiny

Low. This is the same mechanical pattern recently merged for sibling recipes (#1404, #1394, #1416) — image version bumps with a changelog entry. The full-sweep-enabled label ensures the full benchmark sweep will validate it, and the test plan explicitly relies on that.

Other factors

No bugs were found by the bug hunting system. Note the new tag drops the -cu130 suffix that the previous tag had; the rest of the B200/B300 vLLM recipes in this file also use plain v0.21.0 tags, so this aligns with the existing convention rather than diverging from it.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26006166668
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26006166668

[functionstackx]

/reuse-sweep-run

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26019142140
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26019142140