Skip to content

EPIC: Inference dispatch & architecture consolidation (dedicated cycle) #3423

Closed as not planned

Description

@alfredodeza

Tracking epic for the dedicated, non-cadence milestone Inference dispatch & architecture consolidation — everything under #3418's umbrella too large-blast-radius
for a 06x train. Two sub-projects, two specs, one milestone, one epic:

Placement: this milestone has no train-leaves-by deadline (unlike the 06x
cadence). Both sub-projects touch the same forward/dispatch paths that carry
aprender's one proven GPU-parity receipt (Qwen2.5-Coder, docs/BEATS.md) once past
their low-risk Phase 0/1 design work — that's why they're here instead of on 0.69/0.70.

PP-QUANT-001 — quant-type dispatch (#3418, spec: docs/specifications/PP-QUANT-001-MASTER.md)

Phase 0/1 (contract inventory, table + gate design) landed on 0.69.0 via PR #3420 —
low risk, additive, no live call site touched. What lands HERE:

  • Phase 2 — migrate all ~30 dispatch call sites onto quant_type_traits()
  • Phase 3 — parity re-proof against the Qwen2.5-Coder GPU-beat measurement

PP-ARCH-001 — shared-block forward-path consolidation (#3422, spec: docs/specifications/PP-ARCH-001-MASTER.md)

  • Phase 0 — classify every forward function as shared / duplicate / novel
  • Phase 1 — design the composition convention + a slot for new operators
  • Phase 2 — migrate starting from Qwen2.5-Coder, then Qwen3-MoE, then Qwen3.5
  • Phase 3 — prove it on Qwen3-Next / Qwen3.8-Flash landing without the duplicate class

Owner: unassigned.

Activity

  1. added this to the PP-ARCH-001 milestone on Sep 17, 2026
  2. changed the title [-]EPIC: PP-ARCH-001 — shared-block forward-path consolidation (dedicated cycle)[/-] [+]EPIC: Inference dispatch & architecture consolidation (dedicated cycle)[/+] on Sep 17, 2026
  3. noahgift commented on Sep 17, 2026

    @noahgift
    Member

    Operator triage note: #3422, #3423 and PR #3424 were moved into 0.70.0 by this morning's milestone triage run, and have been reverted to this dedicated milestone on reading your body — the no-train-deadline placement stands. No other field was touched.

  4. noahgift commented on Sep 20, 2026

    @noahgift
    Member

    Not scheduled by 30-day-plan.md v2.2 — and that is a decision, not an oversight

    This epic sits on its own milestone ("Inference dispatch & architecture consolidation", 3 open) and the 30-day plan does not allocate a slot to it. Recording why, so nobody reads the silence as an accident.

    Its subject matter is being done, under a different name and a narrower scope. The plan's §5.1 / §5.7 deliver the part of "dispatch consolidation" that a release train can hold:

    What this epic adds beyond those is the wider architectural consolidation (#3422: shared attention/FFN building blocks exist and are 0% adopted). That is real, and it is the kind of work that is invisible at a 72-hour takt: it touches every architecture, it cannot be landed behind one must-carry row, and the plan's own capacity note is that C is ~29 merged PRs per train with C itself [U].

    Consequences under current doctrine:

    • Rule 3 — if this is ever proposed as a must-carry row, it needs a ruled plan and a falsifier proven satisfiable by a RED test on main before that train opens. Today it has neither.
    • Rule 9 — the train exclusion predicates are label-based (arch-model-new, gate-framework-new, …). A consolidation epic is not excluded by them, which means it could drift into a train by default. It should not arrive that way.
    • Its own milestone has no due date, which the plan treats as the correct shape for a bucket that is not a train (cf. 0.73, which is being renamed backlog for exactly this reason).

    Recommendation for the team review: either give this epic a dedicated cycle after the month's spine lands (the plan's spine is E → G → B → A → I, and this is a G/B substrate change), or decompose it into rows that individual trains can carry behind their must-carry. Leaving it as an unscheduled epic with three open children is the state most likely to produce a surprise PR mid-train.

    Plan governance: #3080 · 0.70: #3081.

  5. noahgift commented on Sep 23, 2026

    @noahgift
    Member

    triage 2026-09-23 (aprender-dd, standing triage owner): milestoned to 0.74.0, theme Any Model (architectures, formats); it had no release milestone.

  6. noahgift commented on Sep 25, 2026

    @noahgift
    Member

    Superseded by EPIC E7 #4001 (APR-EPIC-001 A4). Open children re-parented: none. Owner lines copied to the absorbing epic's body. Applied by aprender-d1: claude+agy both said close, apr shadow row recorded (handoff/close-quorum-2026-09-25.json).

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

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions