feat(hybrid): pass token ids to hash-routed layers

[FDecaYed]

What

Teach HybridModel/HybridStack to supply token IDs only to TransformerLayer consumers that use Hash MoE routing.

• pass input_ids through the HybridModel forward path;
• preserve the strict call signature for non-transformer layers;
• keep selective/full recompute and local/TE CUDA-graph execution covered;
• add focused HybridModel coverage without introducing GPTModel support.

This is a child of #5938. Its GitHub Files changed view contains only the HybridModel adapter and its focused test.

Provenance

• Reconstructs the HybridModel integration portion of [dev] [DeepSeek-v4] Part 2: Hash MoE and SwiGLU clamp #4481.
• Refactors the same frozen feature set captured in [Tracking] DeepSeek-v4 HybridModel merge plan #5795.
• Original Hash routing implementation by the [dev] [DeepSeek-v4] Part 2: Hash MoE and SwiGLU clamp #4481 authors; DSv4 adaptation by @hxbai.
• Reconstructed against the feat(moe): add hash-based token routing #5938 merge mirror; feature cutoff: 2026-07-21.
• GPTModel integration is intentionally excluded.

Testing

• isort --check-only, ruff, Black --check, Python compilation, and git diff --check passed for the slice.
• Focused CUDA/Torch execution is delegated to PR CI because this local macOS host has no CUDA devices.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[FDecaYed]

/ok to test 6f8a121

[FDecaYed]

Closing this stacked adapter because it is absorbed into P4: one complete HybridModel-only Hash MoE feature rebuilt through #5938 after P2 merges. Tracking PR: #5795.