Test AIPerf empty-content TTFT on MiniMax TRT

[cquil11]

Summary

• Pin utils/aiperf to the exact head of SemiAnalysisAI/agentx-harness PR [NVIDIA] Update B200 SGL FP8 and add FP4 TP4 #42 (e8ece5d29bd0285af1ced36fbc4c670328b2d58c).
• Enable --allow-empty-content only for the MiniMax-M3 NVFP4 B300 TensorRT-LLM AgentX recipe.
• Use the existing TP8, concurrency-1 configuration to test whether the parser-suppressed first generation event establishes TTFT.

Why

The existing MiniMax-M3 TRT-LLM configuration uses stream_interval: 100. Its first generated reasoning delimiter is suppressed by the reasoning parser, so current AIPerf waits for later non-empty content before recording TTFT. This PR tests AIPerf PR #42 against that concrete serving path.

Validation

• bash -n benchmarks/single_node/agentic/minimaxm3_fp4_b300_trt_mtp.sh
• Focused AIPerf tests: 26 passed
• Matrix generation selects exactly MiniMax-M3 TRT, B300, TP8, concurrency 1.
• A one-off AgentX fast run will be linked after dispatch.

[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]

Claude finished @cquil11's task in 3m 49s —— View job

---

PR Review

• Gather context (diff, changed files)
• Verify perf-changelog entry placement and rules
• Verify aiperf submodule pin (reachability, diff vs current pin)
• Review benchmark script change
• Post review feedback

Summary

Reviewed all 3 changed files. The benchmark script change is sound: REPLAY_CMD is a plain string built via += appends in benchmarks/benchmark_lib.sh:1998, and appending --allow-empty-content after build_replay_cmd (benchmarks/single_node/agentic/minimaxm3_fp4_b300_trt_mtp.sh:176) matches the established pattern other agentic recipes use for flags like --apply-chat-template and --server-metrics. The EVAL_ONLY path is correctly unaffected. The perf-changelog.yaml entry is correctly appended at the end of the file.

One 🟡 warning posted inline on utils/aiperf: the submodule is shared by every agentic recipe, and it's being pinned to an unmerged PR head. Please confirm (1) e8ece5d29 is a fast-forward descendant of the current pin 754356e9a with all behavior changes gated behind the opt-in --allow-empty-content flag — otherwise this silently changes aiperf behavior for all other agentic sweeps while the changelog scopes it to MiniMax only; and (2) the commit is reachable from the configured submodule remote SemiAnalysisAI/aiperf.git (the PR description references SemiAnalysisAI/agentx-harness PR #42, a different repo name). Plan to re-pin to the merged mainline aiperf commit once PR #42 lands. I could not verify the ancestry or the flag's existence at the pinned commit directly (no network access to the aiperf repo from this environment); the queued sweep run will exercise both.

No blocking issues found.

[cquil11]

+viz @ajcasagrande

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32305145240
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32305145240

[claude]

🟡 WARNING: Shared submodule pinned to an unmerged PR head

Why it matters: utils/aiperf is consumed by every agentic recipe (all of them build REPLAY_CMD from the same submodule), not just minimaxm3-fp4-b300-trtllm-agentic-mtp. Two things can't be verified from this diff alone:

1. Ancestry: if e8ece5d29 is not a fast-forward descendant of the current pin 754356e9a (e.g. aiperf PR [NVIDIA] Update B200 SGL FP8 and add FP4 TP4 #42 is based on an older or diverged base), this pin silently changes or rolls back aiperf behavior for all other agentic sweeps, while the perf-changelog entry scopes the change to the MiniMax config only. It's only safe if every behavior change in the 754356e9a..e8ece5d29 range is gated behind the opt-in --allow-empty-content flag.
2. Reachability: the submodule remote is SemiAnalysisAI/aiperf.git, but the PR description says the commit is the head of SemiAnalysisAI/agentx-harness PR [NVIDIA] Update B200 SGL FP8 and add FP4 TP4 #42 — a different repo name. If the commit only exists outside the aiperf repo network, git submodule update will fail on fresh checkouts. Additionally, once PR [NVIDIA] Update B200 SGL FP8 and add FP4 TP4 #42 is rebased or squash-merged, this dangling commit stays reachable only through the PR ref, which is fragile for long-term reproducibility on main.

Fix: Confirm in the aiperf repo that git merge-base --is-ancestor 754356e9a e8ece5d29 holds and that the range is fully gated behind --allow-empty-content; then re-pin to the merged mainline aiperf commit once PR #42 lands, before (or promptly after) this PR merges to main.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32305345287
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32305345287

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32310661720
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32310661720

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32310661720
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32310661720

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32310661720
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32310661720