feat: accept GLM-5.2 AgentX results / 支持接收 GLM-5.2 AgentX 结果

[JordanNanos]

Summary

• Register glm5.2 and zai-org/GLM-5.2-FP8 as a distinct database bucket while grouping the bucket under the existing GLM-5 dashboard model.
• Rename dashboard and comparison labels to GLM-5/5.1/5.2 while preserving the stable glm-5-1 comparison slug.
• Add regression coverage for the H200 + SGLang FP8 + AgentX + DRAM HiCache unofficial-run path from InferenceX run 29648586946.
• Works for both official ingested results and ?unofficialrun= overlays; overlay normalization is covered by the API route test.

Deployment order

Deploy these mappings before ingesting run 29648586946. If ingestion happens first, reingest the run after deployment so glm5.2 rows are retained.

Testing

• pnpm fmt
• pnpm lint
• pnpm typecheck
• pnpm test:unit (3,056 passed)
• pnpm test:e2e: component suite passed (182/182); DB-backed browser specs could not complete because the local database is behind the repository schema (benchmark_type and offload_mode columns are missing).

中文说明

概要

• 将 glm5.2 和 zai-org/GLM-5.2-FP8 注册为独立数据库分桶，同时归入现有 GLM-5 仪表板模型。
• 将仪表板和对比页标签更新为 GLM-5/5.1/5.2，并保留稳定的 glm-5-1 对比页 slug。
• 针对 InferenceX 运行 29648586946 的 H200 + SGLang FP8 + AgentX + DRAM HiCache 非官方运行路径补充回归测试。
• 官方入库结果和 ?unofficialrun= 叠加层均受支持；API route 测试已覆盖叠加层归一化逻辑。

部署顺序

请先部署本 PR 的映射，再摄取运行 29648586946。如果已经提前摄取，需要在部署后重新摄取，确保 glm5.2 数据行不会被跳过。

测试

• pnpm fmt
• pnpm lint
• pnpm typecheck
• pnpm test:unit（3,056 项通过）
• pnpm test:e2e：component suite 全部通过（182/182）；由于本地数据库落后于仓库 schema，缺少 benchmark_type 和 offload_mode 列，依赖数据库的 browser specs 无法完成。

---

Note

Low Risk
Additive model mappings and label updates with broad unit test coverage; deploy mappings before ingesting GLM-5.2 runs to avoid dropped rows.

Overview
Adds GLM-5.2 as its own DB bucket (glm5.2, HF zai-org/GLM-5.2-FP8) while keeping it grouped under the existing GLM-5 dashboard model, following the same point-release pattern as glm5 / glm5.1.

Mappings and ingest: DB_MODEL_TO_DISPLAY / inverse grouping now include glm5.2; ETL MODEL_TO_KEY and resolveModelKey resolve the glm5.2 prefix and HF path. Compare pages keep the stable glm-5-1 slug but query glm5.2 plus older buckets and show GLM 5/5.1/5.2 in headers and dropdown labels.

Regression test: Unofficial-run normalization is covered for the H200 + SGLang FP8 AgentX + DRAM HiCache overlay path (cluster:h200-dgxc, kv_offload_backend: hicache). AGENTS.md documents GLM-5.2 parameter counts.

^{Reviewed by Cursor Bugbot for commit 47422ba. 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, Comment	Jul 18, 2026 3:07pm

[JordanNanos]

Closing because the GLM-5.2 ingestion support was superseded by merged PRs #588, #589, and #591. Any follow-up to regroup GLM-5/5.1/5.2 should be based on the current master branch.

中文说明

由于 GLM-5.2 数据摄取支持已由合并的 PR #588、#589 和 #591 取代，现关闭此 PR。后续如需将 GLM-5/5.1/5.2 重新归为同一组，应基于当前 master 分支进行修改。