Skip to content

V3 bar vs code gap: apr serve runs Qwen3.5 one request at a time, so S3 cannot pass #5046

Closed as not planned

Description

@noahgift

What

The adopted V3 bar needs m_formed == c (S3) at c = 4/8/16 on the blessed Qwen3.5-4B. But apr serve serves the hybrid model one request at a time by design (#3571): the app state holds a single Mutex<Qwen35Session>, and the code calls the hybrid "a single-stream model". There is no batch scheduler, so no Batch m=N done lines are produced. S3 cannot pass on the blessed model until a batched hybrid serving path exists. S7 may still pass, since each request decodes alone once it holds the session.

Done when

Either (a) a batched Qwen3.5 serve path exists and S3 passes at c = 4/8/16, or (b) the bar is changed by a ruling. For (b), the owner runs a planted quorum on bar versus batched path itself (C314), and a bar change needs a non-author sign-off round.

Proposed by la-0.71.1a (inbox 08:14:16Z). Refs #4971.

Activity

  1. added this to the 0.71.1 milestone on Oct 10, 2026
  2. noahgift commented on Oct 14, 2026

    @noahgift
    MemberAuthor

    merged into #5312 (master ticket for epic #3994; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    owner:la-0.71.1aOwned by look-ahead session la-0.71.1a

    Type

    No type

    Projects

    No projects

      Milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions