Repository navigation
E7: qwen3_moe model-family contract first (contract → code → parity) #5056
Description
Child of #4001 (Any Model).
First architecture added contract-first: contracts/model-families/qwen3_moe.yaml, validated by pv, with falsification tests, plus the MoE rows in apr-architecture-schema-v1, arch-constraints-v1 and architecture-requirements-v1. Code is derived from the contract afterwards (GH-323 codegen path), never written ahead of it.
Done when: pv validate is green on the family contract; ArchConstraints for qwen3moe is generated from it; CPU parity vs llama.cpp is measured on more than one prompt. CUDA MoE parity waits on #4378.
Activity
feat(contracts): qwen3_moe model-family contract first, with falsifie…
feat(format): qwen3_moe reaches codegen + Architecture::Qwen3Moe (Refs …
docs(lookahead): 0.74 kit R1 written, falsifiers green (Refs #5056)
docs(lookahead): 0.74 kit R5 done_when amendment drafted (Refs #5056)
fix(serve): arch_requirements codegen reads the in-tree contract (Refs …
feat(contracts): is_moe axis in architecture-requirements, MoE cells …
fix(serve): arch_requirements fallback snapshot carries the is_moe ce…
docs(lookahead): 0.74 kit R4 written, falsifiers green (Refs #5056)
noahgift commentedon Oct 10, 2026
Finding rows (la-0.74, on la-74/wip, no PR; these ride this ticket's PR, and no new ticket is minted because the E7 child budget is 3/3):
crates/aprender-core/build.rsreadscrates/aprender-core/contracts/model-families/(16 families, fix: PMAT-517 — publish safety gate prevents broken crates.io releases (19 incidents) #701) before the rootcontracts/model-families/(29). As a result, 13 root families never reachbuild_default_registry, andqwen3_5.yamlalready differs between the two copies. qwen3_moe is guarded by FALSIFY-MF-QWEN3MOE-005 (the two copies must be byte-identical); no other family has a guard.cuda/types.rsWeightValidation skipped the dense FFN whenis_moe. This is fixed here: the contract now has theis_moecells (1.1.0) and the skip is deleted. (b)aprender-core/src/format/layout_contract_enforce.rskeeps its own hand-written(qk_norm, bias)table. That table has no MoE row and is still open. (c)arch_requirements_fallback.rsis regenerated in this branch. Also, KANI-ARCH-001..003, which the contract names, have no Rust harness.build.rslooked for../../../provable-contracts/, which resolved only from~/src/aprender, to the archived sibling repo. Every worktree and CI build therefore used the fallback.build.rsemits norerun-if-changedfor it, so restoring the file does not regenerate the code. This was observed on intel: a planted mutant ran against a stale fallback untilbuild.rswas touched. This is still open.R4 evidence: on intel @7459baae12, FALSIFY-ARCH-013 is green both through the contract and through the fallback (54/54 each). A planted mutant (the MoE cell changed to
base) turns 3 of the 54 tests red.fix(serve): a missing arch-requirements contract watches its parent d…
noahgift commentedon Oct 10, 2026
F4 fixed on la-74/wip @84160930ff: a missing architecture-requirements YAML now watches its parent dir. Intel lane (move YAML away, build, restore without touching build.rs, build): control 8f388bd stays on the fallback, 8416093 regenerates from the contract. A second no-op check with the YAML missing reruns nothing (#4614 cost not reintroduced).
docs(lookahead): 0.74 kit staged head 8416093, F4 fixed (Refs #5056)
test(core): FALSIFY-MF-014, deleting any required family field turns …
70 remaining items
docs(lookahead): A1 step 27, the MoE expert matvec runs the one table…
refactor(serve): fused_matmul keeps no list of fused types; the table…
docs(lookahead): A1 step 28, fused_matmul keeps no list of fused type…
docs(lookahead): Q2 answered (a), the A5 parity pass line is a 32-tok…
docs(lookahead): A5 scored on the Q2 line from existing receipts, goo…
docs(lookahead): A5 passes the Q2 line on all 16 prompts, good 16 of …
docs(lookahead): A5 near-tie check, llama.cpp's own top-2 margin at t…
docs(lookahead): A5 print fix staged on la-74/print-ids, so apr run -…
docs(lookahead): Q3 answered (a), A4 passes when a new arch adds no p…
docs(lookahead): A4 shape test staged on la-74/a4-forward-allowlist, …
docs(lookahead): C358 1c re-run, all 32 dense rows comparable now tha…
noahgift commentedon Oct 14, 2026
merged into #5315 (master ticket for epic #4001; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.
fix(serve): apr run -v prints every encoded prompt id, not the first …