Add Qwen3.5 H200 MTP config / 新增 Qwen3.5 H200 MTP 配置 - #2592
Add Qwen3.5 H200 MTP config / 新增 Qwen3.5 H200 MTP 配置#2592hshrivastava-droid wants to merge 7 commits into
Conversation
添加 H200 MTP 配置。
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
1 similar comment
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
补充性能变更日志中的 PR 链接。
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| - agentic-coding | ||
| description: | ||
| - "Add Qwen3.5 FP8 H200 SGLang AgentX MTP points with DRAM HiCache" | ||
| - "Use TP8/EP1 concurrency 2 through 24, published FlashInfer 0.6.15 CUDA 13 wheels, and golden MTP acceptance length 3.39" |
There was a problem hiding this comment.
🟡 The new perf-changelog.yaml entry sets pr-link to https://github.com/SemiAnalysisAI/InferenceX/pull/XXX instead of the actual PR number. docs/configuration-procedures.md only permits the literal placeholder pr-link: TBD before a PR exists, to be replaced with the real URL immediately after PR creation — since this PR already exists as #2592, the entry should read pull/2592.
Extended reasoning...
The new tail entry appended to perf-changelog.yaml (line 5929) ends with:
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXdocs/configuration-procedures.md documents the changelog-append procedure precisely: the template at line 251 shows pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/<number>, and line 254 explicitly states: "Before the PR exists, the model+hardware playbook permits pr-link: TBD. Replace it with the real URL immediately after creating the PR." The only sanctioned placeholder is the literal string TBD — not a URL with a literal XXX substituted for the PR number.
Why this isn't caught elsewhere: nothing in the repo's YAML schema or CI validates that pr-link matches pull/\d+, so a stray XXX placeholder parses fine as valid YAML and doesn't fail any automated check. The only enforcement mechanism is the documented convention plus the stop-condition in the same doc (line 272): "...or still has TBD when the PR is otherwise ready for sweep." That stop condition is phrased around the literal TBD placeholder, so a pull/XXX link technically slips past it too, even though it's just as unresolved as TBD would be.
Step-by-step proof:
- Every other entry in
perf-changelog.yaml(e.g., lines for PRs perf(qwen3.5-fp4-b200-sglang-mtp): add TEP2 arm on NVFP4-V2 / perf(qwen3.5-fp4-b200-sglang-mtp): 基于 NVFP4-V2 添加 TEP2 组合 #2550, Add DSV4 B200 disaggregated Dynamo SGLang MTP configuration / 新增 DSV4 B200 分离式 Dynamo SGLang MTP 配置 #2554, [AMD] [AgentX] Add MI355X Qwen3.5 FP4 SGLang MTP #2562, [AMD] [AGENTX] GLM-5.2 MTP5 tune: raise acc-len and drop hicache at low conc / GLM-5.2 MTP5 调优：提升接受率目标并在低并发下关闭 hicache #2570) uses a real numeric PR link:.../pull/2550,.../pull/2554, etc. - This PR's own metadata identifies it as PR Add Qwen3.5 H200 MTP config / 新增 Qwen3.5 H200 MTP 配置 #2592, meaning the real PR number was known/knowable at diff-authoring time.
- The correct final value per the documented procedure is
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2592. - As written,
pull/XXXis neither the permitted pre-PR placeholder (TBD) nor a resolved, valid link — it's an artifact left over from the changelog template that was never filled in.
Impact: This breaks changelog traceability — anyone (or any tooling) later using pr-link to trace a config change back to its originating PR will hit a dead/invalid URL (.../pull/XXX) instead of the real PR. It doesn't fail the benchmark sweep or any executable check, so it's not merge-blocking, but it should be corrected before merge since the PR number is already known.
Fix: Change line 5929 from pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX to pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2592.
This reverts commit 9faefbb.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31731134658 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31731134658 |
| @@ -0,0 +1,160 @@ | |||
| #!/usr/bin/env bash | |||
| set -euo pipefail | |||
| set -x | |||
There was a problem hiding this comment.
dont we already have an qwen3.5_fp8_h200_mtp.sh file, can u edit that instead of creating an new one?
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31745690810 |
更新 SGLang 镜像。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31747034666 |
更新 FlashInfer 版本。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31748413253 |
2 participants
Description
中文说明
Related Issue
N/A
Type of Change
Checklist
perf-changelog.yamland have not edited historical entriesOWNER/MEMBER/COLLABORATOR) has commented/reuse-sweep-runon this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the sweep label will no longer automatically kick off new sweeps. Remove and re-add the label to force one.