chore(db): purge run 34512984806 — GB300 DSv4.1 Flash vLLM server got 1 CPU core / 清除 run 34512984806：GB300 vLLM server 仅分配 1 个 CPU 核心

[functionstackx]

Adds run 34512984806 to PURGED_RUNS so it is skipped on ingest and deleted from the DB.

Reason: wrong results. This is the dsv41flash-fp4-gb300-vllm-agentic-dspark AgentX sweep for InferenceX #2961 (Add GB300 DeepSeek V4.1 Flash AgentX). The srun launch command did not pass --cpus-per-task, and the GB300 Slurm cluster default allocated only 1 CPU core to the entire vLLM server, so the ingested throughput/interactivity points are not representative. Root cause is an infra launch bug on our side, not a vLLM or NVIDIA recipe issue. Fixed by InferenceX #3017 (--cpus-per-task=144); the config will be re-run once that lands and the GB300 cluster is out of maintenance.

Ran oxfmt --check and oxlint on the changed file locally (clean). CI applies the override to production, then verifies, invalidates, and warms the cache.

中文说明

将 run 34512984806 加入 PURGED_RUNS，ingest 时跳过并从数据库中删除。

原因：结果不正确。 该 run 是 InferenceX #2961（新增 GB300 DeepSeek V4.1 Flash AgentX）的 dsv41flash-fp4-gb300-vllm-agentic-dspark AgentX sweep。srun 启动命令缺少 --cpus-per-task，GB300 Slurm 集群默认只给整个 vLLM server 分配了 1 个 CPU 核心，因此已入库的吞吐量/交互性数据点不具代表性。根因是我们自身的基础设施启动 bug，与 vLLM 或 NVIDIA 配方无关。InferenceX #3017 已通过 --cpus-per-task=144 修复；待其合并且 GB300 集群维护结束后重新运行该配置。

本地已对改动文件运行 oxfmt --check 与 oxlint（均通过）。CI 会将此覆盖应用到生产环境，随后进行数据库验证、缓存失效与缓存预热。

---

Note

Low Risk
Single audited purge entry in the ingest override ledger; removes bad benchmark data without changing application logic.

Overview
Adds GitHub Actions run 34512984806 to PURGED_RUNS in run-overrides.ts so the ETL pipeline skips ingest for that workflow and removes any already-stored benchmark rows when overrides are applied (CI on merge).

The run is the dsv41flash-fp4-gb300-vllm-agentic-dspark AgentX sweep tied to PR #2961. Metrics are not representative: the vLLM server was started with srun without --cpus-per-task, so Slurm defaulted to one CPU core for the whole server and throughput/interactivity collapsed. That is documented as an infra launch bug (fix in #3017 with --cpus-per-task=144); a re-run is expected after that lands.

^{Reviewed by Cursor Bugbot for commit 19a15e1. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
inferencemax-app	Building	Preview	Sep 11, 2026 5:48pm UTC