Captured-pair gates: score fresh runs against an oracle-side-only top-K golden instead of the self-anchor
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 5/5
- 预计耗时
- 一周以上
- 新手友好度
- 35/100
- Issue 类型
- 功能
- 描述清晰度
- 基本清楚
- 活跃度
- 活跃
- 技术栈
- bash, cpp
- 领域
- backend, testing-qa
调研方向
从 .agents/specs/tenstorrent-keepquant.md 和 scripts/qwen35-q4km-neartie-gap.sh 开始，然后审查 verified-pair 和 #2959 流程。使用独立的 spec 定义 checker 更改，演示一次 red-before ULP re-roll，并提供 mutation 证据，证明真实的 divergence 在 review 之前仍然会失败。
由索引模型根据 Issue 内容生成。
描述
Row: BACKEND-TENSTORRENT-KEEPQUANT
Problem
The captured-pair gates (the qwen3.5 Q4_K_M TT vehicle today; the same
ladder shape in the qwen3/mistral gates) pin two functions of OUR engine's
numerics: the exact ids anchor (self-match REQUIRE) and a per-capture
teacher-forced gap golden. Every ULP-level shift — a tt-metal pin bump, a
kernel change — re-rolls the ~51/256 near-tie cells and turns the gate red,
forcing a golden refresh. The refresh itself is where the process failed on
2026-09-12: an off-recipe re-derivation wrote f32-grain gaps that
manufactured four band violations and nearly drove a kNearTieMnats
widening (recorded in .agents/specs/tenstorrent-keepquant.md, Evidence).
The pinned-recipe truth was max gap 250 mnats, 0 cells above band — the
engine was fine; the refresh path was not.
Proposed shape
Commit an oracle-side-only golden: per prompt cell, the pinned oracle's
top-K logprobs under the ORACLE's own greedy prefix — stable under our code
changes, so it never re-captures for engine-side reasons. The gate scores
the fresh engine run against the committed top-K up to the first
divergence and fails only a confident divergence (gap above band, or our
token outside top-K). Near-tie flips pass with a recorded margin. The
self-anchor exact REQUIRE becomes an advisory MESSAGE recording the
re-roll. At the decision cell this is exactly the bar the gap script's own
header declares ("is our token, given our exact prefix, one the oracle
cannot confidently separate"); post-flip cascade cells are cross-context
and uninformative today anyway.
Why not now
The verified pair plus the provenance-hardened re-derivation path
(scripts/qwen35-q4km-neartie-gap.sh) lands in #2959's flow. This
redesign is a semantic checker change: it needs its own spec, a
red-before demonstration (a ULP re-roll failing today's anchor),
mutation evidence that a real divergence still fails, and fresh review.
- 主要语言
- C++
- 星标
- 423
- 派生
- 53
- 平均合并
- 1 天 7 小时
- 30 天内合并 PR
- 382
环境准备
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
mudler/vllm.cpp 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 84/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 82/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 70/100
维护者通常 1 天内回复
-
难度 1/5 1 小时以内 新手友好度 88/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 84/100
维护者通常 1 天内回复
相似的 Issue
-
area/actorsystem bug tsan
难度 2/5 1-3 小时 新手友好度 74/100
ydb-platform/ydb#54282 ·
维护者通常 1 天内回复
-
bug needs triage
难度 1/5 1 小时以内 新手友好度 90/100
project-chip/connectedhomeip#74434 ·
维护者通常 1 天内回复
-
难度 1/5 1 小时以内 新手友好度 88/100
tenstorrent/tt-metal#58057 · 1 条评论 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 84/100
maplibre/maplibre-native#4690 ·
维护者通常 1 天内回复
-
comp-query-execution
难度 2/5 1-3 小时 新手友好度 84/100
ClickHouse/ClickHouse#122569 ·
维护者通常 1 天内回复