Add DeepSeek V4 architecture adapter

[abhinav-bellapu]

Description

Adds a TransformerBridge architecture adapter for DeepseekV4ForCausalLM, addressing #1466.

DeepSeek V4 does not have a conventional single residual stream or V2/V3-style MLA. It keeps hc_mult residual streams alive between blocks through manifold-constrained Hyper-Connections (mHC), and interleaves sliding attention with Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). This adapter delegates those model-specific calculations to Transformers for parity, while exposing the internal surfaces needed for interpretability:

• mHC stream inputs, collapse weights, stream-mix matrices, and collapsed attention/MLP inputs
• Q/KV and grouped output projections
• CSA/HCA compressed KV states and causal block bias
• Lightning Indexer selections and projections on CSA layers
• hash and learned top-k MoE routers, routed experts, and shared experts
• final hyper-head collapse, RMSNorm, embeddings, and unembedding

The block bridge intentionally does not register conventional residual aliases: block-boundary hooks carry [batch, pos, hc_mult, d_model], so presenting them as an ordinary three-dimensional residual would make patching code silently target the wrong tensor. Collapsed sublayer inputs remain directly patchable at blocks.{i}.attn_hc.hook_out and blocks.{i}.mlp_hc.hook_out.

The architecture is registered in the factory and all model-registry sites, and V4-specific compression, mHC, indexer, and router metadata is preserved on bridge.cfg for analysis tools.

Validation

The integration suite builds a download-free ~40K-parameter DeepseekV4ForCausalLM spanning all three attention layer types and both router types.

• Exact bridge-vs-Hugging Face FP32 logit parity (atol=0, rtol=0)
• mHC stream/collapse hook shapes and intervention patchability
• CSA/HCA compressed KV and block-bias hooks
• CSA Lightning Indexer output and HCA/sliding optional-component boundaries
• Hash-MoE and learned top-k router/expert hooks
• Focused adapter + registry suite: 15 passed
• DeepSeek V2/V3/V4 integration suite: 34 passed
• Docstring suite: 18 passed, 25 skipped
• mypy .: success across 309 source files
• pycln --check --all, isort --check-only, black --check, and git diff --check: clean

A broad local unit sweep reached 3,215 passed, 34 skipped, 42 deselected, and 9 expected failures. Its remaining 18 setup errors required unavailable Hugging Face downloads, and one unrelated dataset test was blocked by the local sandbox's shared-memory restriction. CI should provide the clean-environment result.

As scoped in #1466, this PR does not claim full Flash/Pro checkpoint verification: the smallest real checkpoint is 284B total parameters and requires cluster hardware. The tiny model exercises the architecture and numerical parity without that hardware dependency.

Type of change

• New feature (non-breaking change which adds functionality)

Checklist

• I have commented non-obvious behavior
• I have made corresponding registry/documentation changes
• My changes generate no new warnings
• I have added tests that prove the adapter and hooks work
• New focused and family tests pass locally
• I have not rewritten tests for existing interfaces