Added Ouro architecture adapter and integration tests for OuroForCausalLM

[Canonik]

Description

Adds a TransformerBridge architecture adapter for OuroForCausalLM (ByteDance Ouro / LoopLM), a looped-depth "Universal Transformer" decoder: one weight-tied stack of 24 physical layers is applied total_ut_steps=4 times per forward pass, with a learned early-exit gate scored alongside the LM head.

What"s included:

• "supported_architectures/ouro.py, including a prepare_loadingpatch for two transformers-v5 incompatibilities in the Ouro remote code (written against 4.55): the removedROPE_INIT_FUNCTIONS["default"]entry, and the missingcompute_default_rope_parametersstatic method that the v5_init_weightsmachinery expects on rotary classes. Both patches are scoped to the importedmodeling_ouro` module; the shared transformers dict is untouched.
• Four-site registration, an "ouro" model_type fallback in sources/transformers.py, and a "ByteDance/Ouro-" entry in _BRIDGE_REMOTE_CODE_PREFIXES in verify_models.py so registry verification auto-enables trust_remote_code.
• A ByteDance/Ouro-1.4B registry row (status 0, null scores).
• Unit adapter test (23 structural tests, runs in CI) and a gated integration test (slow + skipif(CI), QUARANTINES.md row) asserting logit parity against an independently loaded HF eager model at 1e-3, delegation identities, and the per-step hook firing count.

Verified locally on real Ouro-1.4B weights: 11/11 integration tests pass. Loading requires trust_remote_code=True; no new package dependencies.

Note for reviewers: plain AutoModelForCausalLM.from_pretrained on Ouro is currently broken under transformers v5 regardless of TransformerLens (the upstream remote code reads config.pad_token_id, which v5 no longer materializes). The bridge load path already shims this generically; the reference fixture in the integration test mirrors the same shim.

Fixes #1470

Type of change

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

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

[jlarson4]

Incredibly tiny nitpick, apologies, can you add the line number for the test to the link text similar to the other lines in this table?

[jlarson4]

This looks pretty good @Canonik! Just a couple small notes:

1. Missing supports_fold_ln = False – compat mode corrupts the UT loop. Ouro applies model.norm (mapped as ln_final, a weighted RMSNorm) after EVERY UT pass, with the normed output feeding the next pass and the early-exit gate (modeling_ouro.py:592-608), it is not a final-only norm. The adapter leaves supports_fold_ln at the default True, so enable_compatibility_mode() / verify_models Phase 3 folds ln_final.w into W_U, resets the norm weight to ones in the live HF module the loop reuses, and corrupts passes 1..N-1.]. The fix is one line plus a comment: self.supports_fold_ln = False in __init__Related docstring note: ln_final hooks fire total_ut_steps× per forward and run_with_cache keeps only the last pass, worth a comment alongside the existing per-block note.

2. Run verify_models on a sample model, if possible – no models have been verified thus far. The error documented above would've been caught by a verification pass. If you can't run the verification suite on your machine, no worries, just let me know and I can run the verification run & push the results.

3. Surface total_ut_steps (and possibly early_exit_threshold) on bridge.cfg per the config-attr propagation rule. Then the integration test won't have to reach into hf_eager.config for it.

[Canonik]

Thanks I’ll address these in a small follow-up commit.

For verify_models, I’ll try running it locally on the sample Ouro model and report back if I hit an OOM or any local machine limitation. I think I should be able to fit it cleanly, but I’ll confirm with the actual run.

[Canonik]

• 'verify_models' ran clean (Ouro-1.4B, P1-P4 = 100/100/100/95.9).
• I applied the 'supports_fold_ln = False' fix with a unit test covering it.
• Reworked the integration tests as requested: loop config attrs now read off bridge.cfg, plus a propagation test.
• Fixed the QUARANTINES.md line anchor.

Should be ready to merge :)

[jlarson4]

Looks awesome! Thank you @Canonik, merging now

[Canonik]

Glad to hear 😊