[NV] Add H200 DeepSeek-V4-Pro AgentX recipes / [NV] 添加 H200 DeepSeek-V4-Pro AgentX 配方 - #2364

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

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30310586014
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30310586014

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30310920956
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30310920956

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30311003032
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30311003032

/stage-results

@cquil11 staged run 30311003032: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-07-27~r30311003032

This shared staging slot remains available until the next /stage-results request. Staging workflow

Revoking the standing `/reuse-sweep-run` authorization on this PR (removing the bare command comment from 2026-07-28).

A bare `/reuse-sweep-run` is standing rather than one-shot, and it has been silently swallowing sweeps here. The run at the current head, 30505397990, shows the gate emitting skip-pr-sweep: true with reuse-source-run-id empty — the sweep was suppressed outright, not even pinned to a source.

That matters because a real code change landed after the authorization: commit 61a9767f removed import_squash() and its two calls from runners/launch_h200-dgxc-slurm.sh, so the launcher no longer imports the sglang/nginx images and instead assumes pre-staged .sqsh files on the cluster (import_squash occurrences 3 → 0). That is launcher resolution logic on this config's exact path, and no sweep has ever exercised it. Separately, utils/aiperf moved ed2e040d → 381758a8, which changes AgentX pacing/idle semantics for the trace replay.

The last real evidence, 30311003032 (5/5 multi-node agentic), is at 193a0c96 and predates both.

Re-authorize with an explicit run ID once a fresh sweep lands.

@csahithi the AgentX/AIPerf harness has been updated, please merge origin/main into your branch and refresh your submission. Additional tuning may be necessary depending on the config. I apologize for any inconvenience.

This is an automated message.

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30506436017
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30506436017

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30586537651
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30586537651

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30938672335
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30938672335

/stage-results 30938672335

@cquil11 staged run 30938672335: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-04~r30938672335

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

@cquil11 已将运行 30938672335 发布到预发布环境：https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-04~r30938672335

后续的 /stage-results 请求不会移除此运行；再次发布相同的运行 ID 会更新其预发布数据。预发布工作流

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31435999410
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=31435999410

/reuse-sweep-run 31435999410

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. Please link to GitHub Action workflow that shows this. https://github.com/SemiAnalysisAI/InferenceX/actions/runs/31435999410
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this.
• 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 any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.

Additional detail section:

Disclosure — I authored part of what I am signing off. The three most recent commits on this branch are mine, pushed as a maintainer fix rather than by the PR author: e568bd90a (enable SGLang engine metrics and add the required-metric guard), eb92047fa (merge origin/main, resolving the perf-changelog.yaml and launch_h200-dgxc-slurm.sh conflicts), and 4171fc1e9 (pin srt-slurm v1.0.38 for this lane). The submission itself — recipe, master-config entry, and concurrency grid — remains csahithi's work. Flagging this explicitly so the core maintainer can weigh a second reviewer on the launcher and recipe changes.

Validation evidence. Run 31435999410 ran on the exact PR head 4171fc1e9, attempt 1, and settled success with all five multi-node agentic / jobs executed and green at concurrency 1, 2, 4, 8, and 16 — matching the master entry's conc-list: [1, 2, 4, 8, 16] exactly. Wall clock was 1h51m with setup and canary-select both succeeding, so the matrix genuinely executed rather than being skipped under a reuse gate.

What this run fixed, and how it was verified. The previous sweep on this branch produced trace artifacts with no backend engine series, so the published curve rendered incompletely. Two independent causes, both addressed and both confirmed in the run log rather than assumed:

• The aggregated engine never set enable-metrics, unlike the merged GB300 sibling agg-gb300-tp4-mtp-kvoffload.yaml. The engine now launches with --enable-metrics, visible in this run's python3 -m dynamo.sglang command line.
• More subtly, the lane pinned srt-slurm v1.0.10, whose benchmark_stage.py injects AIPERF_SERVER_METRICS_URLS only under isinstance(runner, AIPerfBenchmarkRunner). This recipe is benchmark.type: custom, whose CustomBenchmarkRunner is not that subclass, so AIPerf was never told where to scrape and enable-metrics alone would not have been sufficient. The lane now pins v1.0.38, the same release the GB300 dsv4 AgentX lane runs, confirmed in the log as git clone --branch v1.0.38 --single-branch https://github.com/NVIDIA/srt-slurm.git.

AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:" is now set, so a future regression fails the job loudly instead of silently publishing a partial artifact. Note that the guard's own confirmation line lives in the srt-slurm logs on the cluster rather than the Actions runner log, so the green run is strong but indirect evidence that it passed.

Speculative decoding and acceptance length. MTP via EAGLE with speculative-num-steps: 3, speculative-eagle-topk: 1, speculative-num-draft-tokens: 4. Acceptance is simulated (SGLANG_SIMULATE_ACC_METHOD: match-expected, SGLANG_SIMULATE_ACC_TOKEN_MODE: real-draft-token) with SGLANG_SIMULATE_ACC_LEN: "2.49", which is exactly golden_al_distribution/dsv4_mtp.yaml deepseek-v4-pro.thinking_on[3]. Thinking mode matches the conditions under which that curve was measured (SGLANG_DEFAULT_THINKING: "1", SGLANG_DSV4_REASONING_EFFORT: high; the curve was captured with {"thinking": true, "reasoning_effort": "high"}). No substituted acceptance target. On the chat-template item: this recipe drives the AgentX path (benchmark.type: custom → agentic_srt.sh, AIPerf replaying real agentic-coding trajectories) rather than the random-token benchmark_serving.py path, and the acceptance length is pinned to a golden curve itself measured through the model's chat template, so the AL distribution is aligned to real-world usage.

Evals — not applicable, left unchecked. The dsv4-fp8-h200-dynamo-sglang-agentic-agg entry declares only scenarios: agentic-coding: with no evals key, so search-space-config.agentic_evals is empty and run-sweep.yml's sweep-agentic-evals job cannot be dispatched. There is no eval workflow to link for this submission; the eval lanes are skipped by construction rather than by omission.

Single-node recipe publication — not applicable. The one recipe in this PR is a multi-node srt-slurm recipe under benchmarks/multi_node/srt-slurm-recipes/** (multinode: true), so the single-node vLLM-recipes / SGLang-cookbook publication requirement does not apply.

Model and scenario scope. MODELS.md lists DeepSeek-V4-Pro as active for Agentic coding, with the MTP-only deprecation recorded as not yet enacted. This submission is the agentic MTP arm on the upstream lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45 image.

No engine or serving-stack patching. No .patch, git apply, sed -i, site-packages edit, monkey-patch, or forked engine wheel; the pinned upstream SGLang image runs as shipped. dynamo.install: true with wheel: "1.3.0.dev20260718" installs the Dynamo router, the declared serving frontend for this framework: dynamo-sglang entry and the standard path for every dynamo-* config in the repo. launch_h200-dgxc-slurm.sh clones NVIDIA/srt-slurm@v1.0.38 and overlays this repo's recipe YAML into its recipes/ tree, which is harness configuration and the established pattern for srt-slurm lanes.

Signed: Ankur-singh

✅✅✅ Verdict: PASS ✅✅✅

✅ Check 0 (CODEOWNER): PASS — Ankur-singh is a named owner of configs/nvidia-master.yaml; all other changed paths are catch-all-only, covered by any recognized CODEOWNER.
✅ Check 1 (passing sweep on in-PR commit): PASS — run 31435999410 (attempt 1) succeeded on the exact PR head 4171fc1e9; all five executed multi-node agentic / jobs are green, matching the entry's conc-list: [1, 2, 4, 8, 16] exactly. single-node */ / eval / lanes don't exist for this AgentX-only multi-node submission.
➖ Check 2 (evals pass): N/A — the entry declares only scenarios: agentic-coding: with no eval selection, so no eval lane is generated (all eval jobs skipped by construction, same as the merged GB300 sibling #2157); accuracy fairness is instead enforced via the golden-AL simulation in Check 11. Run image matches the PR config (lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45, visible in the job names).
➖ Check 3 (recipe linked & merged): N/A — exclusively multi-node submission (benchmarks/multi_node/srt-slurm-recipes/**, multinode: true, framework dynamo-sglang); the recipe-link requirement applies to single-node recipes only.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run 31435999410 posted by Ankur-singh (COLLABORATOR), after the earlier standing bare command was explicitly revoked.
✅ Check 5 (latest template): PASS — every current-template item is present; the two unchecked items (evals, single-node recipe link) are explained in the additional detail section and the catch-all reasoning item is checked.
✅ Check 6 (upstream image & ordering): PASS — image is upstream lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45; engine-first ordering satisfied by the existing dsv4-fp8-h200-vllm-agentic (framework vllm, same cluster:h200-dgxc runner and agentic scenario), plus the dsv4-fp8-h200-sglang* entries.
✅ Check 7 (no deprecated models/scenarios): PASS — dsv4 agentic coding is active per MODELS.md; the deprecated arm is agentic non-MTP ("not yet enacted"), and this PR adds the MTP published arm that transition is waiting for.
✅ Check 8 (no architecture hacks): PASS — no --hf-overrides/model-config edits; EAGLE MTP uses DeepSeek-V4's native NextN head, and the SGLANG_OPT_*/HiCache/SWA settings are kernel- and memory-management knobs that don't reduce model FLOPs.
✅ Check 9 (spec-decode via chat template): PASS — the AgentX replay client drives /v1/chat/completions with --endpoint-type chat (benchmarks/benchmark_lib.sh:1929).
✅ Check 10 (no engine patches): PASS — no .patch/git apply/sed -i/site-packages rewrites or forked engine wheels; the dynamo 1.3.0.dev20260718 wheel is the declared router for framework: dynamo-sglang, and cloning NVIDIA/srt-slurm@v1.0.38 + overlaying the recipe YAML is orchestration harness, not engine modification.
✅ Check 11 (golden simulated AL): PASS — aggregated_environment pins SGLANG_SIMULATE_ACC_LEN: "2.49" with match-expected/real-draft-token; 2.49 equals golden_al_distribution/dsv4_mtp.yaml deepseek-v4-pro.thinking_on[3] for the (3,1,4) EAGLE chain, and the thinking mode matches the curve's conditions (SGLANG_DEFAULT_THINKING: "1", reasoning_effort: high). No synthetic-acceptance knobs on any non-agentic config in this PR.

Note: the sign-off discloses that the signer authored the branch's three most recent maintainer-fix commits; flagged here for the core maintainer's awareness per that disclosure.

/stage-results 31435999410

@cquil11 staged run 31435999410: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-10~r31435999410

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

@cquil11 已将运行 31435999410 发布到预发布环境：https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-10~r31435999410

后续的 /stage-results 请求不会移除此运行；再次发布相同的运行 ID 会更新其预发布数据。预发布工作流