Repository navigation
docs: canonical forbidden-optimizations list enforced by sign-off / 文档：建立禁止优化手段的权威列表并在签核中强制执行 - #4510
Closed
Oseltamivir wants to merge 2 commits into
Closed
Oseltamivir wants to merge 2 commits into
Oseltamivir wants to merge 2 commits into
Conversation
…sign-off Collect the forbidden techniques (architecture FLOP reduction, draft precision lowering, engine patching, acceptance gaming, and unevaluated accuracy-trading levers) in one CONTRIBUTING.md section. The sign-off verifier's new Check 8(c) fails any listed technique not already covered by Checks 8, 10, 11, or 13, judged against the default-branch list. AGENTS.md points agents to it before tuning.
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Move both out of the eval-covered levers. Indexer skipping or sharing is forbidden even when the model or pinned engine enables it by default; recipes must disable engine-default sharing explicitly. INT4 quick-reduce is forbidden; INT6 and INT8 stay allowed with eval coverage. Both apply to configs added or changed from now on, without invalidating earlier runs.
6 of 11 tasks
3 of 11 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The rules for what a submission may not do to get faster numbers are spread across the draft-precision section of
CONTRIBUTING.md, the PR review checklist, and the CI sign-off verifier prompt. There is no single list, and AGENTS.md mentions none of it, so an agent tuning a recipe only meets these rules at sign-off, after the sweep has run.CONTRIBUTING.md#forbidden-optimizations(new): the canonical list, with five categories: architecture FLOP reduction, lowering draft precision (linking the existing Draft-model precision section, unchanged), engine patching and the waiver exception, gaming speculative acceptance, and accuracy-trading levers without eval coverage.AGENTS.md: one bullet pointing agents to the list before they tune a recipe, server flags, or an image.Policy change, going forward: two levers move from "allowed with eval coverage" to forbidden:
SGLANG_MINIMAX_M3_INDEX_TOPK_FREQ=2; recipes must set=1explicitly.*QUICK_REDUCE_QUANTIZATION=INT4); INT6 and INT8 stay allowed with evals.Both apply to configs added or changed after merge and do not invalidate earlier runs. 14 active MI355X recipes use INT4 quick-reduce and will fail sign-off when next changed (including image bumps) until it is removed. All other items already existed in the checklist, the draft-precision section, or the verifier. The PR review checklist template is untouched, so Check 5 (latest template) is unaffected.
AI model disclosure
Prepared with claude-opus-5-5 (Claude Code) for analysis and editing.
中文
禁止使用的优化手段目前分散在
CONTRIBUTING.md的草稿模型精度章节、PR 审查清单和 CI 签核验证器提示词中，没有统一列表，AGENTS.md 也未提及，导致智能体只有在签核阶段（扫描已运行之后）才会遇到这些规则。CONTRIBUTING.md#forbidden-optimizations，作为权威列表，涵盖五类：削减模型架构 FLOPs、降低草稿模型精度（链接现有章节，内容不变）、修补推理引擎及豁免例外、操纵推测解码接受长度、以及缺少评测覆盖的精度折衷手段。AGENTS.md新增一条，提示在调整配方、服务参数或镜像前先阅读该列表。政策变更（仅对之后新增或修改的配置生效，不影响既有结果）：索引跳过/共享（包括引擎默认值，如 ROCm 上 SGLang 默认
SGLANG_MINIMAX_M3_INDEX_TOPK_FREQ=2，配方须显式设为1）以及 INT4 量化 all-reduce 改为禁止；INT6/INT8 在有评测覆盖时仍允许。目前有 14 个 MI355X 配方使用 INT4 quick-reduce，下次修改（包括镜像升级）时将无法通过签核。其余条目均已存在于审查清单、草稿精度章节或验证器中；PR 审查清单模板未改动，因此不影响 Check 5。