Admit hybrid models at Ulysses 8 under the exact contract (GDN-CP scan chain + composed admission)

[kiddyboots216]

Exact GDN under Ulysses context parallelism

Stacked on #54. This PR contains production code and conventional tests only.

What changes

Hybrid Qwen3.5-family models were byte-broken under Ulysses sharding because cross-rank GDN state folding, shard-local chunk rephasing, and short-convolution composition did not preserve the Ulysses-1 program. This PR adds:

1. A canonical scan chain that transports the FP32 running-state operands in declared rank order, with receipt validation and deterministic reinjection.
2. A collator contract that aligns document starts, stream ends, and shard cuts to the 64-token chunk grid.
3. Convolution-halo re-entry through the contracted varlen serving kernel using the previous shard's raw post-projection rows.
4. Fail-closed joint admission for qualified Ulysses degrees and the pinned kernel/toolchain fingerprint.

Gates

• State-chain equality at ranks 2, 4, and 8 with vacuity, ordering, digest, and alignment negatives.
• Convolution-halo locality and byte equality.
• Four-rank collective chain mechanics and eight-GPU model/collator integration.
• Hybrid Ulysses-8 hidden-state and FP32 logprob bytes equal to the Ulysses-1 oracle, with both the attention exchange and state chain proven load-bearing.
• Endpoint replay: 1,540/1,540 decision tokens bitwise equal at Ulysses 1 and 8; K3 exactly 0.0.

Scope

Admission remains evidence-scoped to the qualified hybrid family and Ulysses degrees. Missing collator attestation, incompatible topology, or toolchain drift fails closed.

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Superseded by #57, which consolidates the complete exact trainer-serving stack directly onto main.