[functionstackx]

Description

Applies the B200 retune from #3686 to the GB200 DeepSeek-V4.1-Flash vLLM AgentX arm (dsv41flash-fp4-gb200-vllm-agentic-dspark).

• The image moves from nightly cd10ed6f to vllm/vllm-openai:nightly-dev-arm64-cu130-ac9126e58aa7, the arm64 build of the commit that Update B200 DSV41flash vLLM image and retune TP4/DEP2/DEP4 / 更新 B200 DSV41flash vLLM 镜像并调优 TP4/DEP2/DEP4 #3686 (B200) uses and Update GB300 DSV41flash vLLM image and retune TP4/DEP2 / 更新 GB300 DSV41flash vLLM 镜像并调优 TP4/DEP2 #3652 (GB300) shipped.
• TP4 runs at concurrency 1–128 with FlashInfer sparse MLA attention and Engram tables on transparent huge pages.
• DEP2 (TP1 x DP2 + EP2) at concurrency 8–32 and DEP4 (TP1 x DP4 + EP4) at concurrency 64–128 replace TP2. Both use DeepGEMM MegaMoE behind a consistent-hash vLLM Router 0.1.14.
• FlashInfer autotuning is enabled. All points use FULL_AND_PIECEWISE CUDA graphs with capture sizes in multiples of the six-token DSpark verification block, an MXFP4 indexer KV cache with sparse logits, an fp8 KV cache and gpu-memory-utilization 0.97.
• The GB200 header keeps its 4-GPU node and its 5 s × 1440 health check. The engine ready timeout rises from 3600 s to 7200 s to match B200 and GB300.

Validation

• infx/tests/srt_slurm and infx/tests/matrix pass (528 tests).
• infx.matrix.generate test-config emits 12 points (TP4 ×7, DEP2 ×3, DEP4 ×2). Each matches exactly one zip variant, and every zip list has consistent lengths.
• A GPU sweep through the full-sweep-fail-fast label is pending. This PR claims no performance results yet.

AI model disclosure

• Model/version: Claude Opus 5.5 (claude-opus-5-5) in Claude Code; no delegated agents.
• Role: ported the Update B200 DSV41flash vLLM image and retune TP4/DEP2/DEP4 / 更新 B200 DSV41flash vLLM 镜像并调优 TP4/DEP2/DEP4 #3686 recipe, master config, documentation and perf-changelog changes to GB200, ran the local validation and wrote this description.

Related Issue

None. Companion to #3686 (B200).

Type of Change

• Bug fix
• New feature
• Configuration change
• Documentation update
• Other (please describe)

Checklist

• I have completed the AI model disclosure and kept it current
• I have tested my changes locally
• I have updated documentation if necessary
• For every change that can affect benchmark performance and every recipe addition or modification, I have appended a new entry to the physical end of inferencex-e2e/perf-changelog.yaml and have not edited historical entries
• If this PR can affect benchmark performance or adds or modifies a recipe, it carries exactly one primary sweep label (a maintainer applies it on fork PRs): full-sweep-fail-fast (recommended), full-sweep-enabled, or non-canary-full-sweep-enabled. Optional modifiers all-evals, evals-only, and agentx-fast require a primary label; the last two block reuse while applied.
• Before merging via reuse, an authorized maintainer (OWNER/MEMBER/COLLABORATOR) has commented /use <run_id> (or the legacy /reuse-sweep-run) on this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the primary sweep label will no longer automatically kick off new sweeps. Remove and re-add the primary sweep label to force a new sweep.

中文

改动说明

将 #3686 中 B200 的调优移植到 GB200 DeepSeek-V4.1-Flash vLLM AgentX 配置（dsv41flash-fp4-gb200-vllm-agentic-dspark）。

• 镜像从 nightly cd10ed6f 更新为 vllm/vllm-openai:nightly-dev-arm64-cu130-ac9126e58aa7，即 Update B200 DSV41flash vLLM image and retune TP4/DEP2/DEP4 / 更新 B200 DSV41flash vLLM 镜像并调优 TP4/DEP2/DEP4 #3686（B200）所用 commit 的 arm64 构建，也是 Update GB300 DSV41flash vLLM image and retune TP4/DEP2 / 更新 GB300 DSV41flash vLLM 镜像并调优 TP4/DEP2 #3652（GB300）已合入的镜像。
• TP4 覆盖并发 1–128，使用 FlashInfer 稀疏 MLA attention，Engram 表使用透明大页。
• DEP2（TP1 x DP2 + EP2）覆盖并发 8–32，DEP4（TP1 x DP4 + EP4）覆盖并发 64–128，替代 TP2。两者均使用 DeepGEMM MegaMoE，前置一致性哈希 vLLM Router 0.1.14。
• 开启 FlashInfer autotune。所有点使用 FULL_AND_PIECEWISE CUDA graph，捕获尺寸为六 token DSpark 验证块的倍数，使用带稀疏 logits 的 MXFP4 indexer KV cache、fp8 KV cache，并设置 gpu-memory-utilization 0.97。
• GB200 配方头部保留 4 GPU 节点和 5 s × 1440 的健康检查。engine ready 超时从 3600 s 提高到 7200 s，与 B200、GB300 一致。

验证

• infx/tests/srt_slurm 和 infx/tests/matrix 通过（528 个测试）。
• infx.matrix.generate test-config 生成 12 个点（TP4 ×7、DEP2 ×3、DEP4 ×2），每个点唯一匹配一个 zip variant，且各 zip 列表长度一致。
• 通过 full-sweep-fail-fast label 触发 GPU sweep，结果待出，暂不声明性能结果。

AI 模型使用说明

• 模型/版本：Claude Code 中的 Claude Opus 5.5（claude-opus-5-5）；没有委派其他 agent。
• 工作内容：将 Update B200 DSV41flash vLLM image and retune TP4/DEP2/DEP4 / 更新 B200 DSV41flash vLLM 镜像并调优 TP4/DEP2/DEP4 #3686 的配方、master config、文档和 perf-changelog 改动移植到 GB200，运行本地验证，并撰写本说明。

关联 issue

无。与 #3686（B200）配套。

改动类型

配置变更、文档更新。

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[claude]

Findings marked 🟡 are optional suggestions and need no follow-up push.

[claude]

🔴 This new changelog entry's pr-link uses pull/TBD, which will fail CI's perf-changelog validation for every reviewer/CI run on this PR. validate_perf_changelog.py's validate_added_pr_link only accepts the literal XXX placeholder (or the real pull/<pr_number>) via PR_LINK_PLACEHOLDERS/CANONICAL_PR_LINK; pull/TBD matches neither, so it raises ChangelogValidationError("new PR entry must use ... or an XXX placeholder; found ..."). Fix: use https://github.com/SemiAnalysisAI/InferenceX/pull/XXX (the placeholder other entries use, e.g. line 5598) until the real PR number is known.

Why this was flagged

The entry appended at inferencex-e2e/perf-changelog.yaml:9286 sets pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD. validate_added_pr_link in inferencex-e2e/infx/workflows/validate_perf_changelog.py:139-150 checks the link against PR_LINK_PLACEHOLDERS = {"XXX", ".../pull/XXX"} and CANONICAL_PR_LINK (requires pull/\d+); TBD satisfies neither, so validation raises ChangelogValidationError. On the base branch this same validator is satisfied by using the accepted pull/XXX placeholder (precedent at line 5598); this PR substitutes an unrecognized token instead, breaking the automated changelog-append check that gates merge.

Verification: CORRECTION to the candidate's trigger: this does NOT fail "every reviewer/CI run on this PR." The only pull_request-triggered validator never calls validate_added_pr_link, and matrix/validation.py:999 types pr_link as a bare str with no format check, so the PR check-changelog step PASSES with pull/TBD. validate_added_pr_link is only reached at merge time, via canonicalize_appended_links -> compare_entries (prepare_perf_changelog_merge.py:88,519, calling validate_added_pr_link at line 192-193).

[claude]

🟡 (optional) DEP4 (tp:4, ep:4, dp-attn:true, conc-list [64,128]) at nvidia-master.yaml:8449 shares tp and concurrency with the TP4 arm at :8445, so both get the identical exp-name dsv41flash_tp4_conc64_kvnone_spec-mtp for two different deployments. generate.py's _agentic_entries exp-name formula (infx/matrix/generate.py:1188-1192) encodes model, tp, conc, kv-offload and spec-decoding but never ep/dp-attn. This breaks the uniqueness convention documented at nvidia-master.yaml:1544 and makes filter_exp_names raise 'matched multiple rows' for that name. Fix: fold ep/dp-attn into exp-name for any arm that can share tp+conc with another.

Why this was flagged

Trigger: generating the agentic matrix for dsv41flash-fp4-gb200-vllm-agentic-dspark (full-sweep, generate_config_matrix, or test-config --exp-names) via infx/matrix/generate.py's _agentic_entries. The TP4 row at nvidia-master.yaml:8445 (tp:4, conc-list [1,4,8,16,32,64,128]) and the new DEP4 row at :8449 (tp:4, ep:4, dp-attn:true, conc-list [64,128]) both produce exp-name dsv41flash_tp4_conc64_kvnone_spec-mtp (and the conc128 variant) because generate.py:1188-1192 omits ep and dp-attn. On base no two rows for this key share tp+conc: GB300's DEP2 sibling uses tp:2 and the sglang EP arms use a different tp than their TP arm, so exp-names were unique; this is the first collision. filter_exp_names then raises 'Experiment name(s) matched multiple rows' for that name, blocking targeted selection, and no test in infx/tests/matrix checks exp-name uniqueness across search-space rows, so the claimed 528 passing tests do not catch it.

Verification: nit. The exp-name collision is real and reachable, but its serious consequences are prevented downstream, so this is a convention/tooling issue, not a correctness regression. generate.py:1188-1192 builds the exp-name from tp, conc, kv-offload and spec-decoding, omitting ep and dp-attn, so TP4 (:8445) and DEP4 (:8449) both emit dsv41flash_tp4_conc64_kvnone_spec-mtp, violating the uniqueness convention at :1544 and making filter_exp_names raise "matched multiple rows".

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=37076187881

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=37076187881