feat(attention): add SBHD DSv4 hybrid orchestration

[FDecaYed]

What

Enable the SBHD-only DSv4 Hybrid attention variant:

• compose local sliding-window attention with compressed sparse attention;
• add grouped low-rank output projection;
• validate the model variant, attention ratios, and unsupported modes;
• add a backend-explicit module-spec boundary without coupling to GPTModel or models.backends;
• propagate MTP layer identity through existing attention constructors;
• cover construction, QKV, forward/backward, grouped output, RoPE, config, and backend providers.

This PR owns orchestration only. Its Files changed view does not repeat the RoPE or CSA implementations from its parents.

Stack

This PR is stacked on #5947 and transitively #5944.

#5942 is an independent semantic merge prerequisite for correct end-to-end indexer-loss scaling, but is not a Git-base dependency of this slice.

Provenance

• Reconstructs SBHD Hybrid attention behavior from [dev] [DeepSeek-v4] Part 1: Hybrid Attention with CSA and HCA #4458.
• Includes relevant corrections from [dev] [fix] [DeepSeek-v4] fix dense loss and rope type in DSv4 Hybrid Attention #5018.
• MTP layer identity is adapted from [dev] [DeepSeek-v4] Part 3: MTP support with mHC and new mHC contract #4518.
• Saved-output/inverse-RoPE behavior follows Preserve DSA output across fused inverse RoPE #5526.
• Refactors the frozen feature set collected in [Tracking] DeepSeek-v4 HybridModel merge plan #5795.
• Original implementation by the source PR authors; DSv4 integration by @hxbai.
• Reconstructed for current main; feature cutoff: 2026-07-21.
• This is not a mechanical replay or rebase.

Deliberately excluded

• GPTModel and HybridModel wiring;
• fused CSA/DSA kernels;
• packed-sequence/THD and context-parallel support (CP is rejected early);
• HCA, Hash MoE, mHC, and FP8 marker integration.

Testing

• isort, ruff, Black, Python compilation, and git diff --check passed for the exact slice.
• Focused pytest collection is blocked locally because this host lacks Torch/CUDA; PR CI owns execution.

[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 24d4d07

[FDecaYed]

Closing this draft until Round 2. Its logical successor is D2 (SBHD DSv4 orchestration and indexer-loss scaling), rebuilt after D1 and P6 merge and absorbing #5942. Tracking PR: #5795.