fix(etl): map AMD Kimi-K2.7-Code identifiers to kimik2.7-code / 修复:将 AMD Kimi-K2.7-Code 标识映射到 kimik2.7-code

[Oseltamivir]

Summary

AMD AgentX sweeps label the model with the bare prefix kimik2.7 (no -code) and the MXFP4 model path amd/Kimi-K2.7-Code-MXFP4. Neither was registered in PREFIX_ALIASES or MODEL_TO_KEY, so resolveModelKey (packages/db/src/etl/normalizers.ts) returned null and both mapBenchmarkRow and the eval mapper skipped every row as unmappedModel.

Symptom: the unofficial-run overlay for such a run — e.g. runs/29975243963 ([AMD][AgentX] Add Kimi-K2.7 MI355X vLLM agentic MTP) — fetched its artifacts fine but the /api/unofficial-run response contained 0 benchmarks and 0 evaluations, so ?unofficialrun=… rendered nothing. Official ingest would drop the same rows for the same reason.

I confirmed the root cause by running the real resolveModelKey against the exact values in the run's artifacts (infmax_model_prefix: "kimik2.7", model: "amd/Kimi-K2.7-Code-MXFP4", eval model_prefix: "kimik2.7") — all returned null, while the canonical kimik2.7-code / moonshotai/Kimi-K2.7-Code resolved correctly.

Fix

Two mapping additions in normalizers.ts, both folding into the canonical kimik2.7-code DB bucket:

• PREFIX_ALIASES: 'kimik2.7' → 'kimik2.7-code' — also covers precision-suffixed forms like kimik2.7-fp4 via existing PRECISION_SUFFIX stripping.
• MODEL_TO_KEY: 'amd/Kimi-K2.7-Code-MXFP4' → 'kimik2.7-code' — consistent with existing amd/…-MXFP4 entries (amd/GLM-5.1-MXFP4, amd/Llama-3.3-70B-Instruct-MXFP4-Preview).

Tests

Added a regression case to resolveModelKey covering the bare prefix, the precision-suffixed prefix, the eval model_prefix, the AMD MXFP4 path, and the canonical identifiers. pnpm test:unit (db package), pnpm lint, pnpm fmt, pnpm typecheck all pass.

Scope / caveats

• Pure data-mapping change in the db ETL layer — no chart/UI code, so no overlay-path or /zh work applies.
• This unblocks rendering for AMD K2.7-Code runs. It does not make run 29975243963 itself worth ingesting officially: that run's agentic eval gate failed with em_strict: 0.0 on swebench_lite and only 552/1920 requests succeeded — its data is genuinely broken. The mapping fix simply ensures well-formed AMD K2.7-Code runs resolve going forward.

中文说明

AMD AgentX 基准测试(benchmark)的产物(artifact)用裸前缀 kimik2.7(不带 -code)和 MXFP4 模型路径 amd/Kimi-K2.7-Code-MXFP4 标注模型。二者都未在 PREFIX_ALIASES 或 MODEL_TO_KEY 中注册,导致 resolveModelKey(packages/db/src/etl/normalizers.ts)返回 null,mapBenchmarkRow 与评估(evaluation)映射器把每一行都当作未映射模型(unmappedModel)跳过。

现象: 此类运行的非官方运行(unofficial run)叠加层——例如 runs/29975243963——能正常拉取产物,但 /api/unofficial-run 返回 0 条基准、0 条评估,因此 ?unofficialrun=… 页面空白;官方摄取(ingest)也会因同样原因丢弃这些行。我用运行产物中的真实取值调用 resolveModelKey 验证了根因:上述标识均返回 null,而规范形式 kimik2.7-code / moonshotai/Kimi-K2.7-Code 能正确解析。

修复: 在 normalizers.ts 中新增两条映射,统一归入规范的 kimik2.7-code 数据桶:

• PREFIX_ALIASES:'kimik2.7' → 'kimik2.7-code'——经既有 PRECISION_SUFFIX 去后缀后,也覆盖 kimik2.7-fp4 等带精度后缀的形式。
• MODEL_TO_KEY:'amd/Kimi-K2.7-Code-MXFP4' → 'kimik2.7-code'——与既有 amd/…-MXFP4 条目(amd/GLM-5.1-MXFP4、amd/Llama-3.3-70B-Instruct-MXFP4-Preview)保持一致。

测试: 为 resolveModelKey 增加回归用例,覆盖裸前缀、带精度后缀的前缀、评估用 model_prefix、AMD MXFP4 路径以及规范标识。pnpm test:unit(db 包)、pnpm lint、pnpm fmt、pnpm typecheck 全部通过。

范围/注意事项: 这是 db ETL 层的纯数据映射改动,不涉及图表/UI 代码,故不适用叠加层(overlay)路径或 /zh 相关工作。此修复解除了 AMD K2.7-Code 运行的渲染阻塞,但并不意味着 29975243963 这一运行值得官方摄取:该运行的 agentic eval 质量门以 swebench_lite em_strict: 0.0 失败,且仅 552/1920 个请求成功,数据本身有问题。本修复只是保证今后格式正确的 AMD K2.7-Code 运行能够正常解析。

---

Note

Low Risk
Scoped to static model-key aliases in the db ETL normalizers; no auth, API, or UI behavior changes beyond correctly ingesting/rendering rows that were previously skipped.

Overview
Fixes AMD AgentX Kimi-K2.7-Code runs being dropped as unmapped models during benchmark/eval ETL and unofficial-run overlays.

resolveModelKey now maps the bare prefix kimik2.7 (and precision-suffixed variants like kimik2.7-fp4) to the canonical kimik2.7-code bucket via PREFIX_ALIASES, and maps the model path amd/Kimi-K2.7-Code-MXFP4 via MODEL_TO_KEY—matching other AMD MXFP4 entries.

A regression test covers prefix, eval model_prefix, AMD path, and existing canonical identifiers.

^{Reviewed by Cursor Bugbot for commit f876a12. 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 23, 2026 10:39am

[Copilot]

Pull request overview

This PR fixes a db-layer ETL normalization gap where AMD AgentX artifacts label Kimi-K2.7-Code using non-canonical identifiers, causing resolveModelKey to return null and downstream ingest/overlay pipelines to skip all rows as unmappedModel. It updates the shared normalizer mappings so those AMD identifiers fold into the existing canonical kimik2.7-code DB model key and adds a regression test to prevent reintroduction.

Changes:

• Add a PREFIX_ALIASES entry mapping kimik2.7 → kimik2.7-code (including precision-suffixed variants via existing suffix stripping).
• Add a MODEL_TO_KEY entry mapping amd/Kimi-K2.7-Code-MXFP4 → kimik2.7-code.
• Add a regression unit test covering both the bare/precision-suffixed prefixes and the AMD MXFP4 model path.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
packages/db/src/etl/normalizers.ts	Extends model-key normalization with AMD Kimi-K2.7-Code prefix/path aliases to resolve to kimik2.7-code.
packages/db/src/etl/normalizers.test.ts	Adds a regression test ensuring the new AMD identifiers resolve to the canonical key across both benchmark and eval row shapes.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.