Adopt the canonical MoE fold in the Qwen exact trainer combine (fixes trainer/sampler K3 divergence)

[kiddyboots216]

Pair the Qwen exact MoE combine with the serving canonical fold

Stacked on #55. This PR contains production code and conventional tests only.

Root cause

The exact Qwen trainer still reduced exchanged expert contributors with a reverse-rank BF16 chain after serving moved to canonical_moe_fold_v1, a balanced adjacent-pair BF16 tree. Matched per-rank partials were bitwise equal, but the two folds differed in 148,286/296,960 elements at the first MoE block, producing token-level K3 of 1.4e-3.

Fix

• Name and export the trainer's independent transcription of canonical_moe_fold_v1.
• Replace the Qwen trainer chain with the canonical fold after the unchanged contributor exchange.
• Keep transport and backward semantics unchanged.
• Remove stale ordered-combine assumptions from exact-lane validation.

Gates

• Production-path replay: 124,720/124,720 decision tokens bitwise equal; K3 exactly 0.0.
• Trainer fold equals the serving arithmetic on matched partials.
• The canonical fold remains distinguishable from the retired chain at the measured discrimination profile.
• Gradient engagement and existing Qwen exact-model suites pass.

Generality follow-up

Native-EP combine qualification is a per-family registry of admitted EP sizes, with the existing Qwen validator retained as a compatibility shim. Canonical-MoE trainer admission is table-driven by exact-contract family, certified layer count, and topology. Historical GLM-5.2 cases retain identical validation, while unknown families and unqualified sizes fail closed; a new family requires an explicit registry entry plus byte evidence.

[broly-code-security-scanner]

Broly Security Scan

Note

✅ Clean scan
No vulnerabilities detected in this PR.

Note

Re-scan this PR anytime with /broly scan — useful after /broly undismiss, or to refresh findings without a new push.

Broly — SAST (zai-org/GLM-5.2) · Secrets · SCA · IaC · GH Actions · Base Images · Supply Chain Threats · Exploit Chains · Adversarial Verification

We're continuously improving Broly's accuracy and finding quality — your feedback is valuable. False positives, missed findings, bugs, and feature requests all welcome.