Add Shensi and Shensi-VL models

[zongzhilou]

What does this PR do?

Adds two new models sharing a "DeepRecur" design: Shensi (text,
ShensiConfig / ShensiForCausalLM, checkpoint louzongzhi/Shensi-Nano) built on
the DeepSeek-V4 backbone, and Shensi-VL (multimodal, ShensiVlConfig /
ShensiVlForConditionalGeneration, checkpoint louzongzhi/Shensi-VL-Nano) pairing
the Shensi language tower with a Kimi K3 vision tower (MoonshotAI/Kimi-K3).

Shensi (text)

Three changes vs. the vanilla V4 stack:

1. GDAR-XHC — residual streams reworked following xHC
   (arXiv:2607.14530): the residual runs as
   hc_mult=16 parallel streams with per-stream pre/route/post gating, only the
   active streams are updated per step (no Sinkhorn constraint). The Gated Delta
   Attention Residual replaces the remaining mHC machinery: each stream does
   cross-layer attention internally — a delta rule keeps a per-block slot memory
   with decay/erase/write gates, normalized-key retrieval, and softmax routing over
   the slots of previous blocks — instead of stacking an extra attention module on
   top of the hyper-connection.

2. Block Router & Block UniPool — router sharing across a block of layers
   (arXiv:2603.18297, Path-MoE) combined with a
   block-level unified expert pool (arXiv:2605.06665,
   UniPool): the block's write layer owns the top-k router and the expert pool, the
   read layers of the same block share them — one router and one pool per block
   instead of per layer (parameters tied in the state dict). Shared experts are
   removed following arXiv:2605.11689
   (Slicing and Dicing: shared experts barely move the needle).

3. Block aux_loss & erc_loss — kept at block level from an information-theoretic
   angle: the standard load-balancing aux_loss minimizes the mutual information
   between inputs and routed experts (not strictly required once the path constraint
   removes the bias, but kept as a regularizer), while erc_loss
   (arXiv:2512.23447) maximizes the mutual
   information between outputs and routed experts. Together they give routing
   stability and expert specialization, stacking on the specialization already
   enforced by the path constraint for deep cooperation.

Shensi-VL (multimodal)

Shensi language tower + Kimi K3 vision tower, interleaved into DeepRecur blocks
that rely on GDAR:

• cross blocks exchange evidence/guidance between the towers — language queries
  retrieve vision evidence, language guidance is written back to vision;
• the loop block runs a continuous latent reasoning loop (Coconut-style): the
  language stream is refined with damped write-backs, and the loop stops on a
  data-driven criterion — hidden-stream delta, reasoning-state delta and orbit
  recurrence are combined, no hard iteration cap.

Notes

• Docs: docs/source/en/model_doc/shensi.md, shensi_vl.md; tests:
  tests/models/shensi, tests/models/shensi_vl.
• The VL composite owns the tower layers; blocks only reference slices of them, so
  state dicts stay alias-free and save/load round-trips exactly.

Who can review?

• text models: @ArthurZucker @Cyrilvallez @vasqu
• multimodal models: @zucchini-nlp

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 33226796429
Result: failure | Grafana metrics are not available yet.

[github-actions]

Thank you for your contribution 🤗!

CI Security Gate — automatic approval blocked

This PR was not automatically approved for CI because the security gate failed.

Possible reasons:

• The PR touches 50 or more files — only PRs with fewer than 50 changed files are automatically approved
• A changed file is outside the allowed directories (src/, tests/, docs/, utils/), has a disallowed extension (only .py, .txt, .md permitted outside tests/ and docs/), or is not .md/.yml inside docs/ — this covers files the PR deletes or renames, not only the ones it edits
• A new high-severity security issue was detected in the changed Python files (Bandit check)
• The PR touches a path this repository protects from untrusted PRs, such as the file that decides who reviews it — a maintainer must make that change in a separate PR

See the workflow run for the exact violations.

A maintainer can review and manually approve CI if a finding is a false positive.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: auto, shensi, shensi_vl

[github-actions]

Thank you for your contribution 🤗!

CI Security Gate — automatic approval blocked

This PR was not automatically approved for CI because the security gate failed.

Possible reasons:

• The PR touches 50 or more files — only PRs with fewer than 50 changed files are automatically approved
• A changed file is outside the allowed directories (src/, tests/, docs/, utils/), has a disallowed extension (only .py, .txt, .md permitted outside tests/ and docs/), or is not .md/.yml inside docs/ — this covers files the PR deletes or renames, not only the ones it edits
• A new high-severity security issue was detected in the changed Python files (Bandit check)
• The PR touches a path this repository protects from untrusted PRs, such as the file that decides who reviews it — a maintainer must make that change in a separate PR

See the workflow run for the exact violations.

A maintainer can review and manually approve CI if a finding is a false positive.