DeepEP admission for zero-K3 lanes: canonical-combine adapters, receipts, and gates

[kiddyboots216]

Canonical combine: DeepEP admission for exact MoE lanes

Stacked on #53. This PR contains production code and conventional tests only.

What changes

• Add canonical_combine_v1, a transport-independent per-token top-k MoE combine: FP32 FMA in router-slot order, one terminal BF16 round, deterministic trainer-owned backward, and one writer per element.
• Add fail-closed transport admission for contributor coverage, injectivity, transported metadata, and two-sided accounting.
• Adapt DeepEP normal-mode transport by carrying logical slot contributors through dispatch and validating the received layout before canonical combination.
• Preserve transport/backward ownership while separating byte arithmetic from message movement.

Gates

• In-process transport-layout, padding, negative-admission, gradient, and serving-arithmetic identity tests at component and production hidden sizes.
• Four-rank collective transport equality against the canonical oracle.
• Real world-8 DeepEP forward/backward equality across balanced, skewed, and empty-expert routing, including Q397B-class geometry.
• Fused expert-kernel row invariance across batch composition, order, filler rows, and message size.
• Low-latency combine equality with log-formatting disabled, plus world-16 transport validation.

Scope

The combine contract admits only measured transport modes and fails closed elsewhere. Ulysses composition and model-level qualification remain separate downstream gates.

[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.

 

[kiddyboots216]

Superseded by #57, which consolidates the complete exact trainer-serving stack directly onto main.