Add exact deterministic DeepEP training

[kiddyboots216]

Summary

Add the XoRL trainer half of the native deterministic DeepEP zero-K3 contract.

Dependencies:

• Build deterministic DeepEP and SGLang kernel wheels xorl-wheels#1
• Add exact deterministic DeepEP serving xorl-sglang#28

Changes

• add native-exact DeepEP dispatch/combine with one-call deterministic reduction
• support exact Qwen3, Qwen3.5, GLM-5.2, and DeepSeek-V4 model paths
• preserve BF16 communicated expert leaves with the canonical higher-precision fold
• integrate active LoRA, router training/publication, shared-expert joins, packed rows, and context/pipeline parallel layouts
• pin the matching DeepEP and SGLang Kernel release artifacts and the serving submodule commit
• reject unsupported checkpoint, EP-size, and topology combinations before execution

Validation

• focused exact-path trainer tests
• packed-boundary, empty-rank autograd, CP, PP, LoRA, and router-sidecar regressions
• deterministic DeepEP release-lock and ABI checks
• forward/backward and optimizer-update component coverage
• source formatting and static checks

[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.