Repository navigation
PP-QUANT-001 M4: tensor_byte_size and ONE refusal site read TRAITS — closes the #3091-reopen / #3341 class for all 35 ids at once #3432
Description
Parent: #3421 · Slot: 0.69.0 must-carry, after M1. Fixes the #3091 reopen.
Mechanism (measured, a672677f4 + the v0.68.0 release asset): QuantizedGGUFTransformer::tensor_byte_size has 11 arms — no IQ*, and no Q3_K or Q5_1 although the crate defines both. Two more private copies exist (convert/q4k_conversion_stats.rs:6, convert/q4k_converter_helpers.rs:20). Public 338 MB repro: unsloth/Qwen3.5-0.8B-GGUF UD-IQ2_XXS → exit 8, type 16.
Fix: byte_size(t, n) = n / blck_size(t) * type_size(t), precondition n % blck_size == 0, checked arithmetic, from TRAITS. A tensor whose type has no kernel yet is refused by exactly one function that names tensor, type name and id.
Falsifier: byte-size functions keyed on a raw qtype 3 → 1; M3's mixed-qtype fixture loads past tensor_byte_size; contracts/quant-dispatch-v1.yaml ships in the same PR (pv validate 0 failed).
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
noahgift commentedon Sep 17, 2026
Exit fixtures (operator decision 2026-09-17). Two public files from
unsloth/Qwen3.5-0.8B-GGUFat revision6ab461498e2023f6e3c1baea90a8f0fe38ab64d0:apr 0.68.0release assethttps://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/6ab461498e2023f6e3c1baea90a8f0fe38ab64d0/Qwen3.5-0.8B-UD-IQ2_XXS.ggufa369165c8ec45a92d55cbad98b5377e2c32ca8dfc824c899d9d05b33b6b53e54tensor_byte_size· type 16https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/6ab461498e2023f6e3c1baea90a8f0fe38ab64d0/Qwen3.5-0.8B-UD-Q4_K_XL.gguf3177ebd67afe4438374da19e690bc1b98756f7e0fea9240e1be404336156a7b5tensor_byte_size· type 23Provenance: 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, notmain.Exit for this ticket: with M4 merged,
apr run <file> --prompt "2+2=" --max-tokens 8 --no-gpuon both files no longer fails intensor_byte_size. That is the first failure only: what happens after the byte size resolves (a type with no dequant kernel) must be the single typed refusal naming tensor, type name and id — not a second private list. Generating text from these files is not this ticket's exit; it depends on tier-0 dequant (#3441) and the Gated DeltaNet CPU work.alfredodeza commentedon Sep 18, 2026
Pulled forward from 0.69.0 ("must-carry, after M1") to 0.68.2 / P0 (EPIC #3477). This is the concrete fix #3091 (now P0) needs to stop refusing real UD-IQ2_XXS/UD-IQ4_XS Qwen3.5 files — it can no longer wait for the rest of PP-QUANT-001 Phase 1 (#3421) to land first.
noahgift commentedon Sep 18, 2026
Measured 2026-09-18 (main 4a538dd and published 0.68.1, lambda-vector + gx10): after #3114 the loader honours
qwen35, but three callers ofrealizar::gguf::unsupported_architecture_reasonstill refuse it with the pre-fix text —crates/apr-cli/src/commands/qa_capability.rs:98(apr qaCapability Match FAIL, 8/12 gates SKIP),parity_refusal.rs:157,ptx_map.rs:441. "ONE refusal site" must mean these three collapse onto the sameTRAITSpredicate the loader reads, orapr qakeeps failing a modelapr runserves. Falsifier:apr qa Qwen3.5-0.8B-Q4_K_M.ggufmust not print "NEITHER the CPU nor the GPU backend" — add it tocontracts/quant-dispatch-v1.yaml. Priority note: #3091's acceptance is gated on this too (itsapr qacell is red on both hosts).test(quant): the loader and the dispatch must agree about every block…
feat(dogfood): model capability ladder — T-2 proves every (architectu…
fix(ladder): a rung that claims only the CPU is not asked whether the…
PMAT-3477: 0.68.2 — Qwen3 CUDA QK-norm restored in the graph and the …