Repository navigation
[PR] fix(relay): the egress fuse 413 says "declared model window", not "effective" — one word, two numbers - #7123
ranxianglei wants to merge 2 commits into
Conversation
…fective" — one word, two numbers (#5172) The 413 egress-window-fuse message called the UNRESERVED wall (metadata.effectiveContextLimit) "the effective model window", but "effective window" elsewhere ([headroom]/nudge/preflight) means the lower RESERVED judging window. One word, two numbers: the nudge line (judges 792,576) and the 413 (judges 1,048,576) look like a contradiction but are the intended #4205 split (the fuse rides the unreserved wall; gap bodies forward → 400 → refold rescue). The looser fuse bar is intended, not a bug — the defect is the mislabel. Relabel both 413 messages ("declared model window (the unreserved upstream wall — nudge/preflight judge against the lower reserved window, see [headroom])") and add the fuse's caliber to the [headroom] line. Log/observability only; no behavior change (the fuse still judges declaredWindow, nudge/preflight still judge reservedContextLimit). Fixes #5172
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.10821-pr.7123.44c6f7cThis exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr7123.tgz
npm install -g packageInstall instructions are refreshed on each push. |
(ai: vllm-qwen/qwen3.8-27b) Verified PR #7123 against the live branch ( What I checked
Targeted tests (local, on the PR head)
CI (head e84d300)
Conclusion: PR #7123 is correct and ready to merge once the re-run goes green. It fixes the #5172 reader trap (one word, two numbers) with wording only — no behavior change. 结构化八问
判定:PR 正确,可合(等 flake 重跑转绿)。 |
(ai: vllm-qwen/qwen3.8-27b) Correcting my earlier report — the Windows lane is NOT a one-off flake. The re-run (attempt 2) failed again, and the CI's own diagnostic says "deterministic break, not a flake". I dug in; here's the accurate picture. The red is not this PR's
It's a master-level born-red on Windows, hitting many PRs
Mechanism: the #7026 test's 4th case sends real Conclusion: PR #7123 is correct and its own tests pass (5/5). The CI red is an inherited master-level Windows failure in the just-merged #7026 test — not caused by this PR. It's mergeable on its merits; the Windows lane stays red until the #7026 test is fixed on master/ci-fixes (tracked in #7108). 结构化八问
判定:#7123 代码正确、自有测试全绿,可合;CI 红为继承的 master 级 Windows 问题(#7026 测试,今日合入),跟踪在 #7108,非本 PR 应修。 |
…ndows red The only red (test windows-latest/22/5) was inherited: this branch's base 4b77eab predates master's 3186e54 (fix(tests): make the shutdown-signal test portable to Windows). Same signature as #4599 floors 1553/1554/1560. No PR-file changes; targeted local: server-process-listeners 4/4 + PR test files 5/5, typecheck clean.
(ai: vllm-qwen/qwen3.8-27b) Inherited red cleared — no code change needed. The only failing check ( What I did this run:
中文结构化八问
判定:红已清除, |
What
Field data from #5172 (DSH 0.2.2-rc.3 + bili 0.1.215): within the same request, nudge judges against 792,576 (the reserved window) while the egress fuse's 413 message judges against 1,048,576 (the declared/unreserved wall) — a gap of exactly 256,000 (= the max_tokens reservation). The 413 calls 1,048,576 "the effective model window" — but "effective window" in [headroom]/nudge/preflight means the lower RESERVED judging window. One word, two numbers.
Layer judgment
The fuse using the looser unreserved wall is the INTENDED #4205 split (gap bodies forward → upstream 400 → refold rescue), not a bug. The real defect is the 413 mislabeling the unreserved wall as "the effective model window", conflating it with the [headroom] reserved caliber.
Fix
Log/observability only — no behavior change (the fuse still judges
declaredWindow, nudge/preflight still judgereservedContextLimit).Regression
tests/issue4079-egress-window-fuse.test.ts: pins "declared model window" + "unreserved upstream wall" and asserts the ambiguous "effective model window" is gone.tests/issue5191-arm-judge-line.test.ts: pins the new [headroom] fuse clause.This PR fixes the FUSE-path instance of the "effective window" word-referencing-two-numbers family (the floor-3 field data in #5172). The original preflight-deadlock root cause in #5172 remains open and is tracked separately.
Related to #5172