Compare MI300X vLLM NIXL AgentX configurations

[cquil11]

Refresh DeepSeek-V4.1-Flash MXFP4 AgentX on MI300X with the pinned October 8 stock vLLM ROCm nightly and native SRT Slurm. Compare TP8 and packed TP4 aggregation with NIXL P4/D4, P4/3D4, and 2P4/2D4. Runs retain full 1M context, automatic golden DSpark acceptance, fresh servers per concurrency, and all eight or sixteen allocated GPUs in the denominator. Aggregate pools use cache-aware SMG; NIXL pools use session-affinity routing. No engine patches.

The 11-point official full-hour sweep is green and staged. Seven setup-only cache-lock failures passed a failed-only retry. Ingestion verified all 11 points across both attempts, 11 linked traces, and one changelog entry. Evals were skipped as requested.

The two-node disaggregated points add generation-speed/throughput tradeoffs:

• P4/3D4 C32: 11,084 total tokens/s/GPU, 103.20 generation interactivity, 11.91s P90 TTFT.
• 2P4/2D4 C64: 16,062 total tokens/s/GPU, 47.35 generation interactivity, 13.15s P90 TTFT.

At matched generation interactivity, these exceed the published interpolated throughput curve by approximately 75% and 38%. Those are interpolation estimates, not measured matched-latency comparisons. Both pay E2E/TTFT penalties. The new aggregate maximum reaches 19,777 tokens/s/GPU, still 3.5% below the older published maximum; this does not improve the entire published curve.

The accepted sweep retains four scored errors under unchanged gates: one aggregate HTTP write error, one prefill HTTP 500, and two aggregate stream truncations strongly matching SMG's stock 30-minute response deadline. Warmup diagnostics and normal end-of-run cancellations are recorded separately; no timeout workaround was applied.

[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。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=37904289844

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=37904289844

[cquil11]

/stage-results 37881375716

[github-actions]

@cquil11 staged run 37881375716: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-10-09~r37881375716

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

[cquil11]

/stage-results 37904289844

[github-actions]

@cquil11 staged run 37904289844: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-10-09~r37904289844

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

[cquil11]

/reuse-sweep-run 37904289844