feat(format): tensor-layout-v1 12-gate PARTIAL discharge#1384
Closed
noahgift wants to merge 2 commits into
Closed
Conversation
Binds FALSIFY-001..012 from tensor-layout-v1 — the canonical source-of-truth contract referenced from CLAUDE.md (LAYOUT-001/002). - 001: ValidatedEmbedding rejected when > 50% zeros (density) - 002: AprTransformer cannot bypass ValidatedEmbedding (Poka-Yoke) - 003: ValidatedWeight rejects any NaN - 004: 94.5% leading zeros rejected (PMAT-234 spot-check bug) - 005: lm_head shape MUST be [vocab, hidden] - 006: aprender + realizar agree on validation (cross-crate parity) - 007: zero `_ =>` catchall in WeightQuantType dispatch - 008: Q6K through Q4K kernel produces detectable garbage - 009: SafeTensors→APR Q4K ≥ 70% token match vs F32 inference - 010: APR has embedded BPE tokenizer (no sibling tokenizer.json) - 011: APR GPU ≥ 80% of GGUF GPU throughput - 012: SafeTensors GPU ≥ 80% of GGUF GPU throughput ## Five Whys 1. Why does tensor-layout-v1 list 12 falsification IDs without algorithm-level discharge? PMAT lints flagged FALSIFY-001..012 as unbound at PARTIAL_ALGORITHM_LEVEL. 2. Why is this the highest-leverage peripheral binding done so far? tensor-layout-v1 is referenced from CLAUDE.md as the canonical layout source-of-truth and from contracts/tensor-layout-v1.yaml §type_enforcement. Coverage drift here is felt across every load/inference path. 12 gates land in one PR. 3. Why a 70% prefix-match threshold for FALSIFY-009 (not 100%)? Q4K quantization introduces small but bounded drift; per the contract's "coherent output" predicate, exact token-id parity is too strict and would Fail every healthy roundtrip. 70% catches the regression class (garbage like "olumbia+lsi") while letting the drift-only deltas through. 4. Why a strict `_ => Fail` for FALSIFY-007 with zero tolerance? The contract is binary-by-design — even one catchall arm in WeightQuantType dispatch enables the silent wrong-kernel regression class (PMAT-232, ALG-006). Tolerance > 0 would let the regression slip back in via "we'll fix the catchall later." 5. Why model FALSIFY-008's "garbage" as `max_abs > 1e6 OR has_nan`? Healthy quantized matmul outputs are in roughly [-100, 100] range; a wrong-kernel dispatch that doesn't produce magnitude blow-up or NaN means the format-isolation invariant has failed — formats that LOOK compatible-enough to share a kernel are a silent-corruption class. Adds 41 unit tests including a 7-bucket density sweep and a 7-bucket GPU-parity sweep. Realistic-healthy walks the canonical Qwen2.5-Coder-1.5B / RTX 4090 healthy path; pre-fix walks 12 simultaneous regressions (PMAT-234 density, type bypass, NaN leak, transposed lm_head, cross-crate divergence, catchall, format collision, garbage roundtrip, missing tokenizer, GH-87/88 GPU slowdowns). No runtime % shift; algorithm-level coverage advances by 12 gates.
noahgift
force-pushed
the
feat/tensorlayout-001-012-partial-discharge
branch
from
5f72e79 to
9503c72
Compare
Contributor
Author
auto-merge was automatically disabled
May 12, 2026 09:21
Pull request was closed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Binds the canonical layout source-of-truth contract (referenced from CLAUDE.md as LAYOUT-001/002) at
PARTIAL_ALGORITHM_LEVELvia 12 verdict functions. Highest-leverage peripheral binding to date.Gates bound
[vocab, hidden](not transposed)_ =>catchalls in WeightQuantType matchesPinned constants
AC_TL_MAX_ZERO_FRACTION = 0.50AC_TL_SPOT_CHECK_MAX_ZERO_FRACTION = 0.50AC_TL_GPU_PARITY_FLOOR = 0.80Five Whys
See commit message — captures why 70% prefix-match for Q4K roundtrip, why zero catchalls for dispatch, and why max_abs > 1e6 OR NaN models "garbage."
Test plan
cargo test -p aprender-core --lib tensorlayout_001_012— 41 passed🤖 Generated with Claude Code