[Duyi-Wang]

Update the MI355X DeepSeek-V4-Pro-0813 SGLang MoRI disaggregated AgentX image to lmsysorg/sglang-rocm:v0.5.22-rocm10-mi35x-20261028@sha256:44202f87c180a28a6b7570f56dfb37a2c0e90dbec901b16381b87a9d582b1f27. At C1, use P-TP4 / D-TP8 and enable the AMD mono kernel on decode.

Keep DSpark block size 6 (7 draft tokens), 8 request slots per role, decode graph sizes 1–8, FP8 unified KV, and UMBP offload. Other concurrency settings stay unchanged. Append the performance changelog entry.

Companion cookbook update: sgl-project/sglang#46966.

中文

将 MI355X DeepSeek-V4-Pro-0813 SGLang MoRI 分离式 AgentX image 更新为 lmsysorg/sglang-rocm:v0.5.22-rocm10-mi35x-20261028@sha256:44202f87c180a28a6b7570f56dfb37a2c0e90dbec901b16381b87a9d582b1f27。C1 使用 P-TP4 / D-TP8，并在 decode 启用 AMD mono kernel。

保留 DSpark block size 6（7 draft tokens）、每侧 8 个请求槽、decode graph sizes 1–8、FP8 unified KV 和 UMBP offload。其余并发配置保持不变，并追加性能变更记录。

配套 cookbook 更新：https://github.com/sgl-project/sglang/pull/46966。

AI model disclosure

Prepared by OpenAI GPT-6 (Codex): configuration edits, documentation, and checks. The exact model/version identifier was not exposed by the runtime and could not be verified. No delegated agents were used.

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[claude]

Claude finished @Duyi-Wang's task in 37s —— View job

---

LGTM - no blocking issues found

[claude]

I reviewed this PR and found no bugs. All three changed paths have CODEOWNERS, so an AMD config owner should still take a look. One point for that person to confirm: the new decode arg enable-amd-monokernel: true is not how the other SGLang MI355X recipes in this repo turn on the mono kernel. They use monokernel-backend: amd, for example in dsv41flash/sglang/mi355x-fp4-mtp/agentic.yaml. Please confirm that the 20261028 image accepts this flag name.

What was reviewed:

• The master-config tp 4->8 change is on the override_c1 search-space row only. Rows c16, c32 and c48 already use decode TP8, so the change matches them.
• In override_c1, decode gpus: 8 matches tp-size: 8, and prefill stays at TP4, the same as the master row.
• The perf-changelog block was added only at the end of the file, as one block, with the existing blank-line separator kept. The image is pinned by digest in the same tag@ sha256: form that other entries in amd-master.yaml use.

Extended reasoning...

The change bumps the image for one MI355X DSV4 SGLang disaggregated AgentX config, moves the c1 decode to TP8 with a mono-kernel flag, and adds one perf-changelog entry. It touches no security-sensitive surface. amd-master.yaml and the mi355x-* recipe tree are both covered by CODEOWNERS entries, which rules out auto-approval. I can't verify the new flag name enable-amd-monokernel from the repo, because sibling SGLang recipes use monokernel-backend: amd instead.