Fix bug where attributes were being unintentionally dropped

tests/integration/model_bridge/test_cohere_adapter.py

@@ -3,7 +3,8 @@
 Model: trl-internal-testing/tiny-CohereForCausalLM
   - 2 layers, ~8M params, CPU-safe, no gating required
   - tie_word_embeddings=True by default
-  - logit_scale=0.0625 (1/16)
+  - logit_scale=0.125 (canonical Command-R is 0.0625; tiny diverges so
+    regression tests catch silent-fallback bugs in the passthrough)
 
 NOTE: The tiny model has use_qk_norm=False, so QK-norm is not exercised here.
 Cohere's QK-norm is a per-head LayerNorm inside CohereAttention.forward; it is
@@ -96,8 +97,20 @@ def test_cfg_uses_rms_norm_false(self, cohere_bridge: TransformerBridge) -> None
     def test_cfg_logit_scale_is_float(self, cohere_bridge: TransformerBridge) -> None:
         assert isinstance(getattr(cohere_bridge.cfg, "logit_scale"), float)
 
-    def test_cfg_logit_scale_value(self, cohere_bridge: TransformerBridge) -> None:
-        assert getattr(cohere_bridge.cfg, "logit_scale") == pytest.approx(0.0625)
+    def test_cfg_logit_scale_matches_hf(
+        self, cohere_bridge: TransformerBridge, cohere_hf: Any
+    ) -> None:
+        """Regression: logit_scale must propagate from HF (not silently fall back to 0.0625)."""
+        bridge_scale = getattr(cohere_bridge.cfg, "logit_scale")
+        assert bridge_scale == cohere_hf.config.logit_scale
+        # Anchor 0.125 so a passthrough regression that defaults to 0.0625 also trips here.
+        assert bridge_scale == pytest.approx(0.125)
+
+    def test_cfg_rope_parameters_matches_hf(
+        self, cohere_bridge: TransformerBridge, cohere_hf: Any
+    ) -> None:
+        """Regression: rope_parameters must propagate from HF (same passthrough trap as logit_scale)."""
+        assert getattr(cohere_bridge.cfg, "rope_parameters") == cohere_hf.config.rope_parameters
 
 
 # ---------------------------------------------------------------------------

transformer_lens/model_bridge/sources/_bridge_builder.py

@@ -39,6 +39,9 @@
     "chunk_size",
     # Multimodal
     "vision_config",
+    # Cohere
+    "logit_scale",
+    "rope_parameters",
 ]
 
 

transformer_lens/model_bridge/sources/transformers.py

@@ -502,6 +502,9 @@ def boot(
         "chunk_size",
         # Multimodal
         "vision_config",
+        # Cohere
+        "logit_scale",
+        "rope_parameters",
     ]
     for attr in _HF_PASSTHROUGH_ATTRS:
         val = getattr(hf_config, attr, None)