chore(db): purge run 29741710665 (GLM-5.2 B300 SGLang AgentX non-MTP) / 清理 run 29741710665（GLM-5.2 B300 SGLang AgentX 非 MTP 分支）

[functionstackx]

Summary

Adds GitHub workflow run 29741710665 to PURGED_RUNS in packages/db/src/etl/run-overrides.ts, so ingest skips it and db:apply-overrides deletes its rows from production.

What the run is: the Run Sweep for InferenceX#2281 — "Convert GLM-5.2 NVFP4 B300 SGLang high-throughput arm to DEP (EP8), cap conc at 64" (branch glm5.2-b300-dep-throughput-arm, 2026-07-20). It benchmarks glm5.2-fp4-b300-sglang-agentic via benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh — the non-MTP agentic-coding arm, i.e. the one that runs without speculative decoding.

Why — the AgentX spec-decode policy

Per MODELS.md:

Going forward we no longer benchmark non-spec-decode versus spec-decode as an A/B. The non-spec-decode arm existed as a neutral baseline back when acceptance length wasn't standardized. That is now solved: golden_al_distribution/ commits one golden acceptance-length curve per model, thinking mode, and draft length, measured on the SPEED-Bench coding category, and AgentX pins every submission to that curve through synthetic acceptance (vLLM synthetic_acceptance_length, SGLang SGLANG_SIMULATE_ACC_LEN, TensorRT-LLM TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS, etc). With a fair, engine-independent acceptance target in place, spec-decode results are directly comparable on their own and a separate non-spec-decode track is redundant. Agentic coding recipes are therefore run and published with speculative decoding enabled only — MTP, EAGLE/EAGLE3, DSpark, or whatever draft method the model ships — and the non-spec-decode arm is neither run nor published. New models are onboarded that way from day 0, as Kimi-K3 is.

MODELS.md's deprecation table lists GLM-5.2 explicitly: deprecated arm Agentic coding, non-MTP; published arm Agentic coding, MTP.

Why it is safe to enact now. MODELS.md gated this retirement on the replacement existing — "Removing them today would leave MiniMax-M3 and GLM-5.2 with no active config at all." That blocker is cleared: glm5.2-fp4-b300-sglang-agentic-mtp landed upstream in InferenceX#2447 (merged 2026-08-07), so GLM-5.2 B300 keeps an active agentic config after this purge.

This follows the precedent set by 30405836523 (Kimi-K3 B300 AgentX non-DSpark), the entry directly above the tail of the list.

Tests

packages/db/src/etl/run-overrides.test.ts asserts structural invariants over the registries (positive integers, no overlap between PURGED_RUNS / PURGED_RUN_ATTEMPTS / CONCLUSION_OVERRIDES / PURGED_BENCHMARK_POINTS) rather than per-run membership, so a new entry needs no test change and is covered by the existing suite. 16/16 pass, plus typecheck, lint, and fmt clean.

The entry is inserted in ascending run-id order between 29702212452 and 29811350508, matching the file's existing convention.

中文说明

将 GitHub workflow run 29741710665 加入 packages/db/src/etl/run-overrides.ts 的 PURGED_RUNS，使 ingest 跳过该 run，并由 db:apply-overrides 从生产数据库删除其记录。

该 run 是什么： InferenceX#2281 的 Run Sweep——"将 GLM-5.2 NVFP4 B300 SGLang 高吞吐分支改为 DEP(EP8) 并将并发上限设为 64"（分支 glm5.2-b300-dep-throughput-arm，2026-07-20）。它通过 benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh 测试 glm5.2-fp4-b300-sglang-agentic，即不启用投机解码的非 MTP 智能体编码分支。

原因——AgentX 投机解码政策： 根据 MODELS.md，我们不再以 A/B 方式对比"非投机解码"与"投机解码"。非投机解码分支原本是在接受长度（acceptance length）尚未标准化时的中性基线；现在 golden_al_distribution/ 为每个模型、思考模式与草稿长度提交一条基于 SPEED-Bench coding 类别测得的黄金接受长度曲线，AgentX 通过合成接受率（vLLM synthetic_acceptance_length、SGLang SGLANG_SIMULATE_ACC_LEN、TensorRT-LLM TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS 等）将每份提交都对齐到该曲线。有了公平且与引擎无关的接受率目标后，投机解码结果本身即可直接横向对比，单独的非投机解码轨道已属冗余。因此智能体编码配方一律仅在启用投机解码（MTP、EAGLE/EAGLE3、DSpark 或该模型自带的草稿方法）的情况下运行与发布，非投机解码分支既不运行也不发布；新模型从第 0 天起即按此方式接入，Kimi-K3 便是如此。

为何现在可以执行： MODELS.md 原本要求先有替代分支才能下线——否则会使 GLM-5.2 完全没有可用配置。该前置条件现已满足：glm5.2-fp4-b300-sglang-agentic-mtp 已通过 InferenceX#2447（2026-08-07 合入）进入上游，因此本次清理后 GLM-5.2 B300 仍保有可用的智能体配置。本次改动沿用 30405836523（Kimi-K3 B300 AgentX 非 DSpark）的先例。

测试： run-overrides.test.ts 校验的是各注册表的结构性不变量（正整数、四个注册表之间互不重叠），而非逐个 run 的成员关系，因此新增条目无需改动测试，现有用例即可覆盖。16/16 通过，typecheck、lint、fmt 均通过。新条目按文件既有约定，以 run id 升序插入于 29702212452 与 29811350508 之间。

🤖 Generated with Claude Code

---

Note

Low Risk
Single registry entry for intentional data removal; no ingest logic changes and the published MTP replacement config is already in place.

Overview
Adds GitHub Actions run 29741710665 to PURGED_RUNS in run-overrides.ts (sorted by run id), so ingest skips it and db:apply-overrides / post-merge CI remove any existing benchmark rows for that sweep.

The run is the non-MTP glm5.2-fp4-b300-sglang-agentic AgentX sweep from PR #2281; it is retired under the AgentX spec-decode policy in MODELS.md (agentic coding is published MTP-only), with glm5.2-fp4-b300-sglang-agentic-mtp already landed in #2447.

^{Reviewed by Cursor Bugbot for commit 1631d74. 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 (UTC)
inferencemax-app	Ready	Preview	Aug 7, 2026 8:05pm

[blacksmith-sh]

Found 1 test failure on Blacksmith runners:

Failure

Test	View Logs
Overview page/switches between B200 and 30-day comparison without losing page state	View Logs

^{Need help on this PR? Tag @codesmith-bot with what you need.}