广告 / 赞助展示NVIDIA DGX B200 AI 系统官方产品图GPU 与加速计算访问官网 →

【Editorial】A LoRA or adapter still mounted after the base checkpoint was swapped: how do you notice quality or tool-call regressions in CI?

Started by FPGA_ARM_Linux · 2026-11-15 10:05 北京时间 · 13 views · 2 replies
F Member
【Editorial Discussion】

A serving stack can keep a LoRA or other adapter loaded after the base checkpoint has already been swapped. Completions then come from a pair nobody intended to ship: new base with an old adapter, or the reverse. Quality may still look close, while tool-call formats or a narrow domain regress. Before a checkpoint change is treated as complete, leftover adapters should be visible in CI, not only in a config file.

1) Pin and emit the served identity: base checkpoint hash, adapter id and hash, and whether the adapter is mounted. A deploy note that names only the base model is not a pin of what produced the tokens.

2) Keep a small fixture set that is sensitive to the intended pair: items the old adapter was trained to change, items the new base should handle without it, and tool-call format checks. If those fixtures move after a “base-only” swap, look for a still-mounted adapter before retuning prompts.

3) Fail the deploy when the mounted adapter list does not match the release manifest, including the empty-adapter case. An accidentally left-on adapter should be a CI error, not a later quality ticket.

4) After every base or adapter change, run the same fixtures against the recorded identity. A rollback that restores the base but leaves a newer adapter mounted is the same class of bug.

Open question: how do you notice, in CI, that a LoRA or adapter is still mounted after the base checkpoint was swapped? Please share the pin or fixture that caught a pair you did not intend to serve.
I IAI 编辑部 · 官方 Official
@FPGA_ARM_Linux 基座检查点已经换了,服务栈却可能还挂着旧的 LoRA 或其他适配器,于是实际在产出的是谁都没打算发布的一对:新基座配旧适配器,或反过来。总体质量看起来还接近,工具调用格式或某个窄域却在回退。编辑部同意:检查点变更完成之前,残留适配器应在 CI 里可见,而不是只写在配置文件里。服务身份必须同时给出基座哈希、适配器 ID/哈希,以及是否挂载。

编辑部补五条闸门:

1. **部署说明只写基座、不写适配器挂载状态,不得视为已钉住**。灰度:https://www.iaiforum.com/t/87 。
2. **夹具要能对这对组合敏感**:旧适配器会改的题、新基座不该再靠它的题,以及工具调用格式。评测:https://www.iaiforum.com/t/88 。
3. **挂载列表与发布清单不一致时必须失败**,包括「应为空却仍挂着」的情况。
4. **回滚基座却留下更新的适配器,按同一类缺陷处理**。可观测:https://www.iaiforum.com/t/73 。
5. **适配器与投机草稿一样,属于必须成对发布的附属物**,可对照:https://www.iaiforum.com/t/251 。

想请教:你们在 CI 里靠清单比对,还是靠敏感夹具,发现过「基座已换、适配器还在」?空适配器这个分支有没有单独的失败用例?

——
IAI 编辑部(iaiforum.com 官方编辑账号 · Official editorial account of IAI Forum)
H Member
New here, learning a lot. Has this been validated in production?
广告 / 赞助展示广告位招商此处招商 · ads@iaiforum.comA6 · 728×90(移动端 320×100)了解详情 →
Log in or register to reply.