Qwen3.5 text-only TransformerBridge support

[SamuelePunzo]

Description

Adds production-ready, text-only Qwen3.5 support through
TransformerBridge.boot_transformers(...).

The supported public path is:

from transformer_lens.model_bridge import TransformerBridge

bridge = TransformerBridge.boot_transformers("Qwen/Qwen3.5-0.8B")

This PR intentionally keeps the scope narrow:

• Supports dense text-only Qwen3.5 models via Qwen3_5ForCausalLM.
• Routes HF configs with model_type="qwen3_5" and
  model_type="qwen3_5_text" to the text-only adapter.
• Keeps base transformers>=4.56, while adding an optional
  qwen35 = ["transformers>=5.2.0"] install extra for actual Qwen3.5 use.
• Adds clear optional dependency errors when Transformers is too old or does
  not expose Qwen3_5ForCausalLM.
• Preserves text-only routing for top-level multimodal configs by swapping
  config.text_config into the HF causal LM loading path.
• Rejects preloaded full Qwen3_5ForConditionalGeneration / multimodal models
  with a clear text-only error.
• Keeps Qwen3.5 MoE and legacy HookedTransformer.from_pretrained out of
  scope.
• Documents Qwen3.5 hooks under blocks.N.attn.* for full-attention layers
  and blocks.N.linear_attn.* for GatedDeltaNet linear-attention layers.

No linked issue.

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.

Testing

Run in LLMbenchmark-env with the repo's real pytest config:

conda run -n LLMbenchmark-env python -m pytest tests\unit\test_qwen3_5_adapter.py -q

Result:

64 passed

Qwen adjacent regression coverage:

conda run -n LLMbenchmark-env python -m pytest tests\unit\test_qwen3_5_adapter.py tests\unit\test_qwen3_next_adapter.py tests\unit\model_bridge\test_qwen3_moe_adapter.py tests\integration\model_bridge\test_qwen3_moe_bridge.py -q

Result:

135 passed, 2 skipped

Syntax check:

conda run -n LLMbenchmark-env python -m compileall tests\unit\test_qwen3_5_adapter.py transformer_lens\model_bridge\sources\transformers.py transformer_lens\model_bridge\supported_architectures\qwen3_5.py

Full local test-suite status:

• Full collection sees the local suite, but some integration modules download
  HF assets at import time and fail in this environment because socket access
  is blocked.
• tests\unit was run broadly. After the Qwen3.5 test-harness fix, the result
  was 1635 passed, 30 skipped, 10 xfailed, 10 failed, 68 errors.
• The remaining failures/errors are outside this Qwen3.5 surface and are
  dominated by missing network/cache access for HF models such as distilgpt2,
  facebook/hubert-base-ls960, bert-base-cased, facebook/opt-125m,
  openai/gpt-oss-20b, and solu-*, plus a few pre-existing unrelated
  environment/version failures.

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
  • Focused tests emit existing hybrid-layer hook alias warnings for optional
    attention submodules on linear-attention layers.
• 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
  • Focused Qwen3.5 and adjacent Qwen tests pass. The broad local unit suite has
    unrelated environment/network failures as noted above.
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

This is an excellent improvement on the existing Qwen3.5 support, just one small note about package management.

[jlarson4]

@SamuelePunzo Just a heads up, due to some file reorganization the merge conflict on this is pretty complex. I am going to resolve it and push the fix to your branch, make sure you pull it down if you do any additional work on the tests themselves.

[jlarson4]

Hi @SamuelePunzo! The conflict has been resolved and CI is running now. Please pull the latest changes before continuing your work.

Two additional notes:

1. There were additional Qwen 3.5 tests added between when you started this PR, please take a look at those and factor them into your changes
2. Please run make check-format and uv run mypy . & resolve any issues you encounter there to ensure your code can pass CI, those two checks are currently failing.

Thanks again for your hardworking on this, I look forward to getting it incorporated into the next TransformerLens release!

[SamuelePunzo]

Thanks again for the help here! I pulled the latest branch and added packaging>=23.0 to the qwen35 extra as requested.

I also chased down the remaining CI failures after the conflict resolution. The format and mypy checks are passing locally now, and I added a small compatibility fix for the GPT-2 bridge benchmark/generation checks that were failing in CI.

Checks I ran locally:

• make check-format
• uv run mypy .
• uv lock --check
• Qwen3.5 adapter tests: 74 passed
• GPT-2 benchmark check
• generation compatibility tests

GitHub CI is oknow too.

[jlarson4]

Excellent work here! Merging and will include in the next release. Thank you for your contribution @SamuelePunzo