[Proposal] Add a bridge-vs-HF numerical parity test for BERT

[jlarson4]

Proposal

Base on dev-4.x. Add tests/integration/model_bridge/test_bridge_vs_hf_bert_parity.py: a bridge-vs-HF numerical parity test for a BERT checkpoint (e.g. bert-base-cased).

Motivation

The bridge's BERT path has no numerical certification anywhere in CI. The only BERT coverage that exists today is either structural or HookedEncoder-side:

• tests/unit/model_bridge/supported_architectures/test_bert_adapter.py asserts component-mapping shape, weight-conversion patterns, and config flags. It never runs a forward pass.
• tests/acceptance/test_hooked_encoder.py and tests/integration/test_create_hooked_encoder.py certify HookedEncoder

Pitch

• Boot via TransformerBridge.boot_transformers("bert-base-cased", device="cpu", dtype=torch.float32) and run a forward. Boot auto-detects BertForMaskedLM, so no model_class is needed for the MLM case.
• Assert bridge MLM logits match a raw HF BertForMaskedLM forward via torch.testing.assert_close. A local fp32 check on bert-base-cased currently gives a max abs difference of exactly 0.0.
• Don't call enable_compatibility_mode() in the parity test – Compatibility mode applies HookedTransformer's default weight processing (including unembedding centering), which shifts logits away from raw HF.
• Assert run_with_cache activations using bridge cache such as embed.hook_out, blocks.0.hook_resid_pre, blocks.0.attn.hook_q, ln_final.hook_normalized, unembed.hook_out.
• Model on the pythia-70m template; do not straight-copy the HT suite

Acceptance:

• MLM logit parity vs raw HF eager within fp32 tolerance
• ≥1 run_with_cache assertion using correct bridge hook names
• Test lives in tests/integration/model_bridge/, not skipped
• make unit-test passes
• uv run mypy . passes

Checklist

• I have checked that there is no similar issue in the repo (required)

[mohdUwaish59]

Hi @jlarson4 please Can I take this?

[jlarson4]

For sure! You are now assigned @mohdUwaish59

[mohdUwaish59]

Thank you so much, I will work on it now.