Update kimik2.5-fp4-mi355x-vllm vLLM ROCm image to v0.21.0

[Klaud-Cold]

Summary

• Updates the vLLM ROCm image tag for kimik2.5-fp4-mi355x-vllm from v0.18.0 to v0.21.0.

Ref #1154

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]

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.

[claude]

🔴 The new perf-changelog entry sets pr-link: XXX instead of a real PR URL. Every other entry in the file uses the format https://github.com/SemiAnalysisAI/InferenceX/pull/<num>, so this should be https://github.com/SemiAnalysisAI/InferenceX/pull/1407 (or the template form /pull/XXX that utils/merge_with_reuse.sh auto-rewrites). The bare XXX will land in main as a malformed entry and also breaks merge_with_reuse.sh's endswith('/$PR') assertion on the auto-resolution path.

Extended reasoning...

What the bug is

The new entry added to perf-changelog.yaml at lines 2495-2501 sets pr-link: XXX — a bare literal placeholder rather than a GitHub PR URL. Every other entry in the file (100+ entries, e.g. lines 2476, 2482, 2488, 2495) follows the convention pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/<num>, and AGENTS.md (around lines 118-123) documents this format explicitly. The PR description references #1407, so the intended value is https://github.com/SemiAnalysisAI/InferenceX/pull/1407.

Why nothing catches it earlier

The Pydantic schema in utils/matrix_logic/validation.py:483 declares pr_link only as a required str — no URL pattern, no prefix check. So a bare XXX passes schema validation and the PR can land as-is.

Concrete impact on merge_with_reuse.sh

Walking through utils/merge_with_reuse.sh for this exact value:

1. The auto-resolution path inspects the new block's pr-link. With pr-link: XXX, the check "XXX" in link is True, so the block is treated as a PR-side contribution that needs rewriting.
2. The next step does block.replace("/pull/XXX", f"/pull/{pr}") — but the bare string XXX contains no /pull/XXX substring, so the replacement is a no-op and the literal XXX survives unchanged.
3. The subsequent assertion last['pr-link'].endswith(f'/${PR}') then evaluates 'XXX'.endswith('/1407') → False, killing the merge script.

So even ignoring the cosmetic/data-quality issue, this bare placeholder actively breaks the documented auto-rebase tooling.

Downstream tooling

Beyond merge_with_reuse.sh, anything that parses pr-link to render a hyperlink (changelog rendering, dashboards) will produce a broken or non-clickable entry, since XXX is not a URL.

How to fix

Replace the bare XXX with the real PR URL:

  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1407

Alternatively, use the documented template form https://github.com/SemiAnalysisAI/InferenceX/pull/XXX so that merge_with_reuse.sh can substitute the real PR number at merge time.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=25956503208
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=25956503208

[Oseltamivir]

/reuse-sweep-run

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=25976976970
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=25976976970