GLM5.1-FP4-MI355X-SGLang: bump Image to v0.5.12.post1-rocm720-mi35x-20260529

[chunfangamd]

In the current Image lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260529, it fixes the GSM8K accuracy regression reported in sgl-project/sglang#25742 (v0.5.12-20260517 dropped to ~0.32 at TP=2).

Local eval-only runs with this new image recover to gsm8k strict-match 0.975 at TP=2/conc=64 and 0.974 at TP=4/conc=16.

In one AMD CI task, it aims to avoid any similar regression in the future.

Co-authors: @1am9trash @Jacob0226

---

Note

Low Risk
Single benchmark image pin and changelog entry; no sweep or serving-flag changes. Main risk is unvalidated throughput/regression on the new container outside the cited GSM8K evals.

Overview
Bumps the GLM-5.1 MXFP4 MI355X SGLang benchmark (glm5.1-fp4-mi355x-sglang) to lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260529, replacing v0.5.10rc0-rocm720-mi35x-20260415. Model, runner, and fixed-seq-len TP/concurrency sweeps are unchanged.

The image update targets a GSM8K accuracy regression on intermediate SGLang builds (notably ~0.32 strict-match at TP=2 on v0.5.12-20260517, per sgl-project/sglang#25742). Local MI355X eval-only runs on the new tag report ~0.975 strict-match at TP=2/conc=64 and ~0.974 at TP=4/conc=16, above the 0.92 gate in sgl-project/sglang#26396.

perf-changelog.yaml documents the bump and rationale for PR #1593.

^{Reviewed by Cursor Bugbot for commit 857aeea. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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.

[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]

LGTM, straightforward image tag bump to pick up the upstream GSM8K accuracy fix.

Extended reasoning...

Overview

Single-line image tag bump for the glm5.1-fp4-mi355x-sglang AMD CI benchmark entry in .github/configs/amd-master.yaml, moving from v0.5.10rc0-rocm720-mi35x-20260415 to v0.5.12.post1-rocm720-mi35x-20260529. Model, runner, TP/concurrency sweeps, and scenario shapes are unchanged.

Security risks

None. This is a CI benchmark config; the new image tag points at the same lmsysorg/sglang-rocm repository as other entries in the file (which already use v0.5.12.post1 tags for other models), and no application code or permissions are touched.

Level of scrutiny

Minimal — this is a sandbox/CI config change with no production impact. The PR description links to the upstream regression issue (sgl-project/sglang#25742) and reports local GSM8K eval recovery to 0.975/0.974 at TP=2/TP=4.

Other factors

The diff is exactly one line; the surrounding entry is byte-identical otherwise. No CODEOWNER-sensitive paths, no logic changes, no outstanding reviewer comments to address.

[chunfangamd]

/sweep test-config --config-files .github/configs/amd-master.yaml --config-keys glm5.1-fp4-mi355x-sglang

[github-actions]

@chunfangamd Kicking off a sweep.

Run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/26682937987
Command: test-config --config-files .github/configs/amd-master.yaml --config-keys glm5.1-fp4-mi355x-sglang
Pinned ref: 857aeea
Approval: not required (trusted collaborator).