[Klaud Cold] Update qwen3.5-fp8-mi325x-sglang-agentic-mtp SGLang ROCm image to v0.5.19-rocm720-mi30x-20260907

[functionstackx]

Summary

Update SGLang ROCm image for the MI325X Qwen3.5-397B-A17B FP8 AgentX MTP recipe from lmsysorg/sglang:v0.5.16-rocm720-mi30x (v0.5.16 release) to lmsysorg/sglang-rocm:v0.5.19-rocm720-mi30x-20260907 (2026-09-07 ROCm 7.2 mi30x nightly).

• Tag verified on Docker Hub: last pushed 2026-09-07T13:58:10Z, digest sha256:e5c773e5c53bbbb341f20584f92040be17b30f7518a2bbe1781f716cd6ea89e1.
• Same rocm720-mi30x nightly flavor the MI300X Qwen3.5 FP8 AgentX sibling already runs; image repo moves from lmsysorg/sglang (release tags) to lmsysorg/sglang-rocm (nightlies), as that sibling did.
• benchmarks/single_node/agentic/qwen3.5_fp8_mi325x_mtp.sh is unchanged and pins nothing image-specific: AITER attention + unified attention + allreduce fusion, fp8 quantization, fp8_e4m3 KV, native EAGLE MTP, golden AL 3.39. Concurrency grids unchanged.
• Same-day counterpart of the Qwen3.5 SGLang AgentX bumps on B200 ([Klaud Cold] Update qwen3.5-fp8-b200-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260907-30705c00 #2861, [Klaud Cold] Update qwen3.5-fp4-b200-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260907-30705c00 #2862), H200 ([Klaud Cold] Update qwen3.5-fp8-h200-sglang-agentic-mtp (+hicache-mtp) SGLang image to nightly-dev-cu13-20260907-30705c00 #2868), and H100 ([Klaud Cold] Update qwen3.5-fp8-h100-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260907-30705c00 #2869).

Recipes touched: qwen3.5-fp8-mi325x-sglang-agentic-mtp

Test plan

• full-sweep-enabled sweep passes on cluster:mi325x-amds.

🤖 Generated with Claude Code

---

Note

Low Risk
Config-only image pin for a single benchmark recipe; no launch-script or search-space changes, though sweep results may shift with SGLang v0.5.19.

Overview
Bumps the container image for qwen3.5-fp8-mi325x-sglang-agentic-mtp from the release tag lmsysorg/sglang:v0.5.16-rocm720-mi30x to the dated nightly lmsysorg/sglang-rocm:v0.5.19-rocm720-mi30x-20260907, aligning with the MI300X AgentX sibling’s sglang-rocm nightly flavor.

Model, runner, agentic-coding scenario, and TP/EP/MTP concurrency grids are unchanged; only the pinned image moves. perf-changelog.yaml gains an entry for this config key (digest, parity with same-day Qwen3.5 AgentX bumps on NVIDIA).

^{Reviewed by Cursor Bugbot for commit 3729e06. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[claude]

Beyond the inline finding, I also checked two things and ruled them out: the digest sha256:e5c773e5... in the changelog entry is a valid 64-hex-char sha256 (not truncated/malformed), and the recipe has multinode: false with no separate model.container field, so the multi-node "model.container must equal image" invariant doesn't apply here.

Extended reasoning...

This is a two-file, config-only image-bump PR (configs/amd-master.yaml image tag change plus a matching perf-changelog.yaml append). A confirmed finding was already surfaced and will be posted inline: the new changelog entry's pr-link is left as the literal PRLINK_PLACEHOLDER, which does not match the validator's accepted placeholder set (XXX or the real PR URL), so it will fail validate_perf_changelog.py and block merge. Given that finding exists, I stayed silent on approval and instead verified two adjacent concerns from the diff myself: (1) the digest cited in the changelog description is a well-formed 64-character hex sha256 digest, not a corrupted or truncated one; (2) the recipe being bumped has multinode: false and no separate model.container field, so it is not subject to the "model.container must equal image" consistency rule that applies to multi-node srt-slurm recipes — that specific invariant simply doesn't apply to this single-node recipe, ruling out a duplicate-field mismatch as a concern here.

[claude]

🔴 The new entry's pr-link is the literal string PRLINK_PLACEHOLDER, which is not in the changelog gate's allowed placeholder set (only "XXX" or the exact PR URL are accepted), so validate_perf_changelog.py's validate_added_pr_link will reject this PR and block merge/CI. Fix: use the literal placeholder "XXX" or the real canonical URL https://github.com/SemiAnalysisAI/InferenceX/pull/<PR#> before merge.

Extended reasoning...

utils/validate_perf_changelog.py defines PR_LINK_PLACEHOLDERS = {"XXX", ".../pull/XXX"} and CANONICAL_PR_LINK regex; validate_added_pr_link (line ~144-160) raises ChangelogValidationError if link is not in that set and not the exact expected PR URL. "PRLINK_PLACEHOLDER" matches neither, so the changelog gate check fails, blocking the automated merge pipeline for this PR until corrected.

Verification: Severity: normal. The appended perf-changelog.yaml entry ends with pr-link: PRLINK_PLACEHOLDER. The allowed placeholder set is PR_LINK_PLACEHOLDERS = {"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"} (utils/validate_perf_changelog.py:24-27). validate_added_pr_link (lines 144-160) raises ChangelogValidationError on a PR run when the link is not in that set and not the…

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34173811391
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34173811391