Skip to content

PP-QUANT-001 M3: regression fixtures — #1749 #1789 #2535 #3341 (+ #3091 reopen) on tiny synthetic GGUFs, observed RED→GREEN #3429

Closed as not planned

Description

@noahgift

Parent: #3421 · Slot: 0.69.0 must-carry, FIRST in queue.
Mechanism: every one of these was found by a user on a multi-GB download; no fixture in the tree reproduces any of them, so each fix is unfalsifiable after the fact (#3091's single-qtype fixture is how the IQ gap shipped).
Fix: synthetic GGUFs built by the in-tree test factory, tracked with sha256: MoE layer + dense bench path (#1749), MoE + serve route (#1789), complete MoE through validate_quantized_tensors (#2535), Q4_0 experts (#3341), and one mixed-qtype file with an IQ tensor in the shared base (#3091 reopen).
Falsifier: revert each fix commit locally → its fixture goes RED and names its own issue; a fixture that stays GREEN under revert is deleted, not kept.

Ĵ: 24 turns [A] — borrowed; estimate.sh aprender pools 0 of 47 ledger rows today, so there is no measured basis.
Plan + evidence: #3427 · design discussion: #3418 · spec PR: #3420

🤖 Generated with Claude Code

Activity

  1. added this to the 0.69.0 milestone on Sep 17, 2026
  2. noahgift commented on Sep 17, 2026

    @noahgift
    MemberAuthor

    Exit fixtures (operator decision 2026-09-17). Two public files from unsloth/Qwen3.5-0.8B-GGUF at revision 6ab461498e2023f6e3c1baea90a8f0fe38ab64d0:

    file bytes sha256 on apr 0.68.0 release asset
    https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/6ab461498e2023f6e3c1baea90a8f0fe38ab64d0/Qwen3.5-0.8B-UD-IQ2_XXS.gguf 338227456 a369165c8ec45a92d55cbad98b5377e2c32ca8dfc824c899d9d05b33b6b53e54 exit 8 · tensor_byte_size · type 16
    https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/6ab461498e2023f6e3c1baea90a8f0fe38ab64d0/Qwen3.5-0.8B-UD-Q4_K_XL.gguf 558772480 3177ebd67afe4438374da19e690bc1b98756f7e0fea9240e1be404336156a7b5 exit 8 · tensor_byte_size · type 23

    Provenance: bytes and the first 16 hex of each sha256 were measured locally on the downloaded files [V]; the full digests are the Hugging Face LFS oids [X] and agree with the measured prefixes. URLs are pinned to the revision, not main.

    For this ticket: these are too large to track in-tree. The tracked fixture stays a tiny synthetic mixed-qtype GGUF with an IQ tensor in the shared base; these two are the real-file exit check, fetched by sha256 in a non-required lane. Exit: both RED on main today (observed above).

  3. modified the milestones: 0.69.0, 0.70.0, 0.74.0 on Sep 20, 2026
  4. noahgift commented on Sep 23, 2026

    @noahgift
    MemberAuthor

    milestone → 0.74.0 per docs/roadmaps/milestone-triage-2026-09-23.md (#4024), operator approval 2026-09-23 (relayed by aprender-cf). Reason: quant/tensor/arch dispatch consolidation: Any Model core (#4001)

  5. noahgift commented on Sep 25, 2026

    @noahgift
    MemberAuthor

    Parked by APR-EPIC-001: over E7 budget (rank 4/12). Reopen = pull. Applied by aprender-d1: claude+agy both said close, apr shadow row recorded (handoff/close-quorum-2026-09-25.json).

  6. modified the milestones: 0.74.0, backlog on Sep 25, 2026
  7. added
    parkedClosed over epic budget (APR-EPIC-001 rule 5). Reopen = pull.
    on Sep 25, 2026
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

    P1High priorityenhancementNew feature or requestparkedClosed over epic budget (APR-EPIC-001 rule 5). Reopen = pull.

    Type

    No type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions