[AgentX][sglang] DeepSeek-V4 B300 MTP recipe, conc 1 / 新增 DeepSeek-V4 B300 SGLang MTP AgentX 配方（conc 1） - #2452
[AgentX][sglang] DeepSeek-V4 B300 MTP recipe, conc 1 / 新增 DeepSeek-V4 B300 SGLang MTP AgentX 配方（conc 1）#2452yhyang201 wants to merge 4 commits into
Conversation
中文：新增 DSv4 B300 SGLang AgentX MTP 配方。
|
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 |
中文：回填 DSv4 B300 SGLang MTP changelog 条目的 PR 链接。
yhyang201
changed the title
yhyang201
added
the
agentx-fast
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| - "Pin simulated acceptance to the committed golden AL: SGLANG_SIMULATE_ACC_LEN=2.49 (golden_al_distribution/dsv4_mtp.yaml, thinking_on, K=3), SGLANG_SIMULATE_ACC_METHOD=match-expected, SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token; EVAL_ONLY runs keep real verification" | ||
| - "Bump image from lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233 to lmsysorg/sglang:nightly-dev-cu13-20260802-131bd51b, which reads SGLANG_SIMULATE_ACC_TOKEN_MODE (added in v0.5.16)" | ||
| - "Search space is one point: TP8, kv-offloading none, conc 1 — same topology as the STP entry TP8 row and the vLLM dsv4-fp4-b300-vllm-agentic-mtp TP8 arm" | ||
| pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/PENDING |
There was a problem hiding this comment.
🔴 perf-changelog.yaml:5366 sets pr-link to .../pull/PENDING, but this PR is #2452. This isn't just a stale link — utils/validate_perf_changelog.py's validate_added_pr_link() only accepts the canonical pull/<pr_number> link or the XXX placeholder, so PENDING will fail changelog validation on merge/ingest. Please change it to pull/2452 (or XXX if the number wasn't known yet).
Extended reasoning...
The newly appended changelog entry for dsv4-fp4-b300-sglang-agentic-mtp sets:
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/PENDINGEvery other entry in perf-changelog.yaml uses a real numeric PR link (e.g. pull/2213, pull/2312, pull/2360), and AGENTS.md's changelog template only shows pull/XXX as a fill-in-the-blank placeholder — never as a literal committed value. PENDING is neither of those things.
Why this actually breaks validation (not just cosmetic):
utils/validate_perf_changelog.py defines:
CANONICAL_PR_LINK = re.compile(r"https://github\.com/SemiAnalysisAI/InferenceX/pull/\d+")
PR_LINK_PLACEHOLDERS = {"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"}and validate_added_pr_link(link, pr_number):
def validate_added_pr_link(link: str, pr_number: int | None) -> None:
if pr_number is None:
if not CANONICAL_PR_LINK.fullmatch(link):
raise ChangelogValidationError(...)
return
expected = f"https://github.com/SemiAnalysisAI/InferenceX/pull/{pr_number}"
if link not in PR_LINK_PLACEHOLDERS and link != expected:
raise ChangelogValidationError(
f"new PR entry must use {expected!r} or an XXX placeholder; found {link!r}"
)This is invoked from compare_entries() for every newly-added entry, which in turn is called by prepare_perf_changelog_merge.py's canonicalize_appended_links() and resolve_conflict_bytes() — i.e. it runs on both PR-time validation and at merge/ingest time.
Step-by-step proof:
- This PR is [AgentX][sglang] DeepSeek-V4 B300 MTP recipe, conc 1 / 新增 DeepSeek-V4 B300 SGLang MTP AgentX 配方（conc 1） #2452, so on a PR run
pr_number = 2452andexpected = "https://github.com/SemiAnalysisAI/InferenceX/pull/2452". - The committed value is
"https://github.com/SemiAnalysisAI/InferenceX/pull/PENDING". - Check
link not in PR_LINK_PLACEHOLDERS:PENDINGis not"XXX"nor".../pull/XXX"→ condition isTrue. - Check
link != expected:".../pull/PENDING" != ".../pull/2452"→ condition isTrue. - Both conditions are
True, soChangelogValidationErroris raised:"new PR entry must use '.../pull/2452' or an XXX placeholder; found '.../pull/PENDING'". - Separately, on a main-branch run (
pr_number is None),CANONICAL_PR_LINK.fullmatch()requires\d+afterpull/, whichPENDINGalso fails to match.
Either path — PR-time validation or main-branch ingest — raises and blocks the changelog step.
Fix: change line 5366 to pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2452 (the actual PR number), or use the literal XXX placeholder form if the number is genuinely unknown at commit time.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30739285264 |
Enable the FP4 indexer, cache reporting and the golden-AL thinking mode, size the decode CUDA graph to max-running-requests, and add a TP4 arm at conc 1 alongside TP8. 中文：为 DSv4 B300 SGLang MTP 配方启用 FP4 indexer、缓存上报和黄金 AL 对应的 thinking 模式，decode CUDA graph 按 max-running-requests 取值，并在 conc 1 上 新增 TP4 臂。
yhyang201
added
full-sweep-fail-fast-no-canary
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30741264877 |
|
At 中文： |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30741373809 |
… SGLang MTP 中文：为 DSv4 B300 SGLang MTP 配方新增 conc 8 的有/无 HiCache 两个点。
1 participant
Smoke test for a SGLang EAGLE/MTP AgentX recipe on DeepSeek-V4-Pro FP4 B300. One point only: TP8, no KV offload, conc 1.
Acceptance is pinned to the golden AL 2.49 (
golden_al_distribution/dsv4_mtp.yaml, thinking_on, K=3).中文：DeepSeek-V4-Pro FP4 B300 上 SGLang EAGLE/MTP AgentX 配方的冒烟测试，只跑一个点：TP8、无 KV offload、conc 1。接受长度按黄金 AL 2.49 固定。