Add GLM-5 DSA TransformerBridge adapter

[zeotrix]

Adds TransformerBridge support for GlmMoeDsaForCausalLM / GLM-5 DSA.

This adds a GLM-MoE-DSA architecture adapter, including a dedicated DSA attention bridge for MLA-style latent attention plus learned top-k sparse-token selection. The adapter is
registered in the required architecture factory and model-registry surfaces, and tiny-random/glm-5.1 is added and verified in the Bridge registry.

Fixes #1406

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

Screenshots

Not applicable.

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

Validation

Ran:

.venv/bin/pycln --all . --exclude "__init__.py"
.venv/bin/isort --check-only transformer_lens/model_bridge/generalized_components/glm_moe_dsa_attention.py transformer_lens/model_bridge/supported_architectures/glm_moe_dsa.py tests/
unit/model_bridge/supported_architectures/test_glm_moe_dsa_adapter.py tests/integration/model_bridge/test_glm_moe_dsa_adapter.py
.venv/bin/mypy .
.venv/bin/python -m pytest tests/unit/model_bridge/supported_architectures/test_glm_moe_dsa_adapter.py tests/integration/model_bridge/test_glm_moe_dsa_adapter.py tests/unit/tools/
test_model_registry.py -k 'TestGlmMoeDsa or TestGlmMoeDsaBridge or TestRegistrySyncedWithFactory' -q

Result: 11 passed.

Model registry verification:

.venv/bin/python -m transformer_lens.tools.model_registry.verify_models --model tiny-random/glm-5.1

The benchmark reported:

VERIFIED: P1=100.0%, P2=100.0%, P3=85.0%, P4=70.0%

Notes:

- make format and uv run mypy . are blocked locally by a resolver/build issue in the default jupyter dependency group: jupyter-ydoc -> y-py==0.6.2.
- The direct .venv tools above pass.
- verify_models updated supported_models.json and verification_history.json, then exited nonzero during cleanup due to an existing torch.mps.synchronize() call on a machine without an
MPS backend.
- black hung locally even when run only on the touched Python files, so I interrupted it.

[jlarson4]

Thanks for putting this together! A couple small adapter notes:

• Please set supports_fold_ln=False on the adapter. The default for adapters is supports_fold_ln=True. MLA latent projections can't fold, so right now folding is attempted and P3 fails (layer_norm_folding/value_bias_folding/weight_modification). Setting it to false skips those tests and resolves the "verified with failing phase" status.
• test_forward_matches_hf uses seq_len=4 with index_topk=4. The top-k selector picks every token and the DSA mask is a no-op. The test validates dense MLA, but never the sparse path. Please bump the sequence length above index_topk (e.g. seq 8, top-k 4) so the mask actually drops tokens and a sparse-masking bug can be caught.

Also, just a heads up, make sure you pull before you make edits. I merged the latest dev into your branch to ensure you have all the latest changes and to resolve any conflicts with supported_models

[zeotrix]

Thanks @jlarson4, addressed both notes.

• Set supports_fold_ln = False on GlmMoeDsaArchitectureAdapter and added a unit assertion for it.
• Updated the GLM-MoE-DSA integration coverage to use seq_len=8 with index_topk=4, so the DSA sparse path is exercised. I also added cache assertions that hook_topk_indices keeps
  top-k size 4 and that hook_dsa_mask contains both masked -inf entries and unmasked zero entries.

Validation: git diff --check passes locally. Focused pytest is still blocked in my local env by dependency setup issues, but the changes are limited to the requested adapter flag and sparse-path test coverage.

[jlarson4]

Looks great! Merged. Thanks @zeotrix