[AMD] Use HIP Top-K for GLM-5.2 MI355X / 为 GLM-5.2 MI355X 使用 HIP Top-K

[jiejingzhangamd]

Description / 描述

English

This PR switches the GLM-5.2-MXFP4 MI355X SGLang AgentX recipe from fused Top-K v2 to the precompiled HIP Top-K path.

• Change only the runtime environment selection: SGLANG_OPT_USE_TOPK_V2=true → false.
• Keep the image, model, topology, concurrency matrix, HiCache settings, and all other launch arguments unchanged from PR [AMD] Bump GLM-5.2 MI355X SGLang daily image / 更新 GLM-5.2 MI355X SGLang daily 镜像 #3081.
• Append the required performance changelog entry.

At concurrency 8, the supplied A/B result shows a material P90 interactivity gain with essentially unchanged throughput:

DSA backend	Top-K backend	P90 interactivity (tok/s/user)	Throughput per GPU (tok/s)
Triton	Top-K v2	96.97	9,553.2
Triton	HIP Top-K	106.69	9,539.0

HIP Top-K improves P90 interactivity by 10.0% while throughput per GPU changes by -0.15%.

中文

本 PR 将 GLM-5.2-MXFP4 MI355X SGLang AgentX 配置从 fused Top-K v2 切换到预编译 HIP Top-K 路径。

• 仅修改运行时环境变量选择：SGLANG_OPT_USE_TOPK_V2=true → false。
• 保持 PR [AMD] Bump GLM-5.2 MI355X SGLang daily image / 更新 GLM-5.2 MI355X SGLang daily 镜像 #3081 中的镜像、模型、拓扑、并发矩阵、HiCache 设置及其他启动参数不变。
• 按仓库要求在性能 changelog 末尾追加记录。

并发 8 的 A/B 数据显示：HIP Top-K 的 P90 interactivity 提升 10.0%，同时单卡吞吐变化仅为 -0.15%。

Type of Change / 变更类型

• Bug fix
• New feature
• Configuration change
• Documentation update
• Other

Validation / 验证

• bash -n benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
• Parsed configs/amd-master.yaml, configs/runners.yaml, and perf-changelog.yaml with PyYAML.
• Generated the exact glm5.2-fp4-mi355x-sglang-agentic-mtp matrix and confirmed all eight AgentX points are unchanged.
• Validated the append-only changelog contract against origin/main.
• git diff --check
• Full AgentX sweep and eval (CI).

Checklist / 检查清单

• The runtime behavior change is limited to the Top-K backend environment variable. / 运行时行为变更仅限 Top-K backend 环境变量。
• No image, topology, concurrency, HiCache, or launch-argument changes are included. / 未修改镜像、拓扑、并发、HiCache 或其他启动参数。
• A new changelog entry is appended without modifying historical entries. / 已追加新的 changelog 记录，未修改历史记录。

---

Note

Low Risk
Benchmark-only runtime env toggle and changelog; no application logic, auth, or serving topology changes.

Overview
Switches the GLM-5.2 FP4 MI355X SGLang AgentX benchmark from fused Top-K v2 to the precompiled HIP Top-K path by setting SGLANG_OPT_USE_TOPK_V2=false in glm5.2_fp4_mi355x_sglang_mtp.sh. Launch topology, HiCache, concurrency matrix, and other server flags stay the same; only the Top-K backend env var and its comment are updated.

Appends a perf-changelog entry for glm5.2-fp4-mi355x-sglang-agentic-mtp / agentic-coding documenting the tradeoff: at concurrency 8, P90 interactivity +10% (96.97 → 106.69 tok/s/user) with throughput per GPU essentially flat (-0.15%).

^{Reviewed by Cursor Bugbot for commit aac8318. 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]

Looks good, straightforward config change. Reviewed the env-var flip (SGLANG_OPT_USE_TOPK_V2=true → false) and its comment update in the benchmark script, and the new bilingual perf-changelog.yaml entry appended at the tail with matching numbers and a valid config-key/scenario-type. No other launch args, topology, or matrix points are touched.

Extended reasoning...

Overview

This PR touches exactly two files: benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh, where a single environment variable SGLANG_OPT_USE_TOPK_V2 flips from true to false (switching GLM-5.2 on MI355X from the fused Top-K v2 path to the precompiled HIP Top-K path) with an updated explanatory comment, and perf-changelog.yaml, where one new bilingual entry is appended at the physical tail of the file documenting the change with concrete before/after numbers. I confirmed via git diff that this is the entirety of the change — no image, topology, concurrency matrix, or HiCache settings are altered, and the changelog entry is appended (not editing any historical entry, and not byte-shifting earlier content).

Security risks

None. This is a benchmark configuration/environment-variable change and a documentation-style changelog append. There is no auth, crypto, permissions, injection, or data-exposure surface involved.

Level of scrutiny

Low-to-moderate is appropriate: it's a single-line runtime behavior change plus a matching changelog entry, which is exactly the kind of small, self-contained, pattern-following change the approval guidelines describe. I verified the config-key glm5.2-fp4-mi355x-sglang-agentic-mtp exists in configs/amd-master.yaml and that agentic-coding is an established scenario-type already used throughout the changelog, so the new entry is well-formed and consistent with existing entries.

Other factors

The pr-link: TBD placeholder in the new entry initially looked like a defect (every other of 918 entries has a real PR URL), but CONTRIBUTING.md's own documented changelog template uses an equivalent placeholder (pull/XXX), confirming this is normal pre-merge practice since the PR number isn't known before the PR exists — consistent with it being ruled out by the bug hunter as a non-issue rather than reported as a finding. The bug hunter reported no bugs, and the only other candidate raised (single-operating-point validation applied to the full 8-point matrix) is an inherent, previously-established pattern for this kind of kernel-backend env-var change and was already treated as ruled out rather than a live finding.

This review covers commit 34299f7, which is no longer the latest commit on this pull request; later commits are not covered by it.