[AgentX] Mirror GB300 MTP tuning (#2571) to GB200 with 4-GPU->8-GPU workers

[xinli-sw]

Summary

• Mirrors the GB300 AgentX DSV4 MTP tuning from [AgentX] Tune DSV4 GB300 MTP / 调优 DSV4 GB300 MTP #2571 onto GB200, resizing every worker that used 4 GPUs on GB300 to 8 GPUs on GB200 (agg TP4->TP8, disagg prefill DEP4->DEP8); workers already at 8/12/16 GPUs are copied unchanged.
• Adds 5 new recipe files under benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/ and new dsv4-fp4-gb200-dynamo-vllm-agentic-mtp2-{agg,disagg} keys in configs/nvidia-master.yaml, kept separate from the existing (differently tuned) GB200 AgentX MTP recipes/keys to avoid touching them.
• runners/launch_gb200-nv.sh already had the synthetic-acceptance injection, so no runner changes were needed.

Test plan

• changelog validation CI gate passes (perf-changelog entry uses pull/XXX placeholder, update to this PR's number)
• Full-sweep CI job for dsv4-fp4-gb200-dynamo-vllm-agentic-mtp2-agg / -disagg runs green on the gb200-nv cluster

[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]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31893849841
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31893849841

[claude]

🔴 perf-changelog.yaml is documented as append-only and byte-sensitive (AGENTS.md), but this diff rewrites the prior entry's trailing separator line — changing it from two trailing spaces to a plain empty line — instead of leaving it untouched and appending after it. This mutates historical bytes and fails utils/validate_perf_changelog.py's validate_raw_change check (head_raw.startswith(base_raw)), which is enforced in CI by run-sweep.yml's "Validate perf-changelog matrix" step. Fix: restore the original two-space separator line ( \n) after the pull/2571 pr-link and append the new entry after that.

Extended reasoning...

What the bug is: AGENTS.md states the perf-changelog invariant explicitly: "The file is append-only and byte-sensitive. Preserve all existing bytes and separator whitespace, and append only at the tail." This PR's diff to perf-changelog.yaml violates that invariant by modifying an existing byte sequence rather than purely appending.

Where it happens: At the tail of the base file (verified via git show d136fa6:perf-changelog.yaml), the previous entry ends with:

  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2571
  <- this line is two literal spaces, then a newline

The diff shows:

   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2571
-  
+
+- config-keys:
+    ...

i.e. it deletes the two-space separator line and inserts a plain empty line before appending the new entry. That is an edit to a historical byte, not an append.

Why nothing else catches this: The file is YAML, so a two-space line and an empty line are semantically identical to a YAML parser — nothing in normal linting or schema validation would flag this. The invariant is enforced by a dedicated byte-level tool, not by YAML semantics.

Why it matters / the concrete failure mode: utils/validate_perf_changelog.py's validate_raw_change function requires head_raw.startswith(base_raw) whenever changelog entries are appended, raising ChangelogValidationError('appended entries changed historical perf-changelog.yaml bytes; restore the base file byte-for-byte and append at the end') otherwise. Because the base file's byte immediately after pull/2571\n is a space (' ') and the head file's corresponding byte is a newline ('\n'), head_raw.startswith(base_raw) is False, so this raises. This validator runs in CI as the "Validate perf-changelog matrix" step in .github/workflows/run-sweep.yml (with --base-ref origin/main and the PR head as --head-ref), and is also invoked by prepare_perf_changelog_merge.py / utils/merge_with_reuse.sh, the canonical merge path for these PRs. So this isn't a style nit — it's a required, automated gate that will fail as the diff currently stands.

Step-by-step proof:

1. git show d136fa6:perf-changelog.yaml | tail -1 shows the last line is " " (two spaces), confirmed via cat -A showing $ (two spaces before the end-of-line marker).
2. base_raw (bytes of origin/main's perf-changelog.yaml) therefore ends in ...pull/2571\n \n.
3. The PR diff replaces that final \n line with a bare \n, then appends the new entry, so head_raw at that same offset reads ...pull/2571\n\n- config-keys:....
4. head_raw.startswith(base_raw) compares byte-by-byte: both match through pull/2571\n, but the next base byte is ' ' while the corresponding head byte is '\n' — mismatch, so startswith returns False.
5. validate_raw_change raises ChangelogValidationError, which fails the "Validate perf-changelog matrix" CI job.

Fix: Leave the original two-space separator line intact after the pull/2571 line, and append the new - config-keys: entry after it (rather than replacing it with a blank line). This is a trivial one-line whitespace fix.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31904915528
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31904915528

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31907559278
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31907559278

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31920499327
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31920499327

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31923718025
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31923718025

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31907559278
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31907559278

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31951030152
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31951030152

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31965016666
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31965016666

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31965016666
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31965016666

[kedarpotdar-nv]

/reuse-sweep-run 31965016666

[kedarpotdar-nv]

As a PR reviewer and CODEOWNER, I have reviewed this and have:

• Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
• Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
• Verified that this PR has passed PR validation. GitHub Actions run 31965016666
• Verified that this PR passes evals. GitHub Actions run 31965016666
• Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
• For agentic workloads: verified that speculative-decoding configs (EAGLE / MTP / draft models) run with simulated synthetic acceptance, with the acceptance-length value taken from the committed golden AL curve in golden_al_distribution/ for that model, thinking mode, and draft length. A submission may choose any supported draft length, but it may not substitute a different acceptance target.
• Verified against the current MODELS.md that this PR does not submit a deprecated model, scenario, or model-scenario combination.
• Verified that the model architecture isn't changed with benchmark hacks like using --hf-overrides to skipping indexer for every x layers on models that don't natively support this. As a general rule, we won't accept optimizations that reduces the number of model architecture FLOPs. Anything that makes that same computation run faster is fair game; FLOPs at lower precisions is fine, given that the config passes private evals. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
• If an company claims that they support vLLM/SGLang as first class LLM inference engines on their hardware, I have verified that the respective vLLM submission made using upstream https://hub.docker.com/u/vllm docker repo, upstream SGLang https://hub.docker.com/u/lmsysorg docker repo. The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet as supported by vLLM/SGLang community maintainers
• If an company claims that they support vLLM/SGLang as first class upstream in-tree LLM inference engines on their hardware, I have have verified that the respective vLLM/SGLang submission has been made before additional frameworks (TRT-LLM, ATOM, etc.). The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet.
• Verified that every single-node vLLM/SGLang recipe in this PR is documented in the official vLLM recipes and/or the SGLang cookbook:
  • I linked the corresponding upstream PR in the vLLM recipe repo or SGLang repo and verified that it is MERGED before this InferenceX PR merges. An opened, draft, or closed-without-merge upstream PR does not satisfy this requirement. If the matching recipe was already published, I linked the published recipe/cookbook page in the additional detail section below.
• Verified that this PR does not patch the inference engine or serving stack — the pinned image must run as shipped. This covers .patch files / git apply / patch, inline patches embedded in benchmark scripts (e.g. a python3/sed heredoc that rewrites installed engine sources before serving), in-place edits of site-packages, monkey-patching, overwriting container files, and installing forked/rebuilt engine wheels on top of the pinned image. The only exception is a patch covered by a filled-out waiver at docs/waiver/<PR_NUMBER>.md — named after the PR that introduces the patch and filed in that same PR, stating what is patched, why the unmodified upstream image cannot run this benchmark, the upstream PR/issue link, and the removal plan — which I have linked below in the additional detail section.
• If this PR uses append-only: true, verified that it only adds generated points or recipe variants inside a selected existing config/scenario and existing same-image visual curve: every previously generated point remains present with the same recipe, no prior point is removed or rerun, and every benchmark-affecting change in the complete diff can affect only the corresponding newly appended points (never an existing point), regardless of which file contains it.
• If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.

Additional detail section:

• Run 31965016666 passed all seven unique GB200 AgentX benchmark points and all four selected eval variants.
• GSM8K strict-match scores were 0.9227–0.9295, above the required 0.91 threshold.
• MTP uses three speculative tokens with synthetic acceptance length 2.49, matching the committed DeepSeek-V4-Pro thinking_on golden AL. EVAL_ONLY keeps real target verification.
• AgentX traffic uses the chat-completions endpoint.
• The submission uses the upstream vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f image.
• This PR contains only multi-node recipes, so the single-node upstream-recipe documentation requirement is not applicable.
• The diff contains no engine patch, model-architecture FLOP-reduction override, or append-only: true.
• Sweep reuse was authorized in this comment.

Signed: @kedarpotdar-nv

[cquil11]

/stage-results 31965016666

[github-actions]

@cquil11 staged run 31965016666: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-16~r31965016666

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow