[model] fix: preserve Step3.7 projector bias

Signed-off-by: Yu Yao <yaoyu.094@gmail.com>

Author: yaoyu-33

src/megatron/bridge/models/stepfun/step37_bridge.py

@@ -119,6 +119,9 @@ def megatron_to_hf_config(cls, provider: Step37ModelProvider) -> dict:
         mtp_num_layers = hf_config.pop("num_nextn_predict_layers", None)
         if mtp_num_layers is not None:
             hf_config["text_config"] = {"num_nextn_predict_layers": mtp_num_layers}
+        projector_bias = getattr(provider, "projector_bias", None)
+        if projector_bias is not None:
+            hf_config["projector_bias"] = projector_bias
         return hf_config
 
     def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> Step37ModelProvider:
@@ -535,6 +538,7 @@ def mapping_registry(self) -> MegatronMappingRegistry:
             # inside ``image_insert_embedding.align_projector``; on the HF side
             # it is the top-level ``vit_large_projector`` linear.
             "image_insert_embedding.align_projector.weight": "vit_large_projector.weight",
+            "image_insert_embedding.align_projector.bias": "vit_large_projector.bias",
         }
         for megatron_param, hf_param in vision_replicated_param_mappings.items():
             mapping_list.append(ReplicatedMapping(megatron_param=megatron_param, hf_param=hf_param))

tests/unit_tests/models/stepfun/test_step37_bridge.py

@@ -17,10 +17,13 @@
 from types import SimpleNamespace
 from unittest.mock import patch
 
+from torch import nn
+
 from megatron.bridge.models.conversion.utils import conform_config_to_reference
 from megatron.bridge.models.stepfun import step37_bridge as _step37_bridge_mod
 from megatron.bridge.models.stepfun.configuration_step37 import Step37Config
 from megatron.bridge.models.stepfun.modelling_step37 import transformer_block as _step37_block_mod
+from megatron.bridge.models.stepfun.modelling_step37.image_insert_embedding import ImageInsertEmbedding
 from megatron.bridge.models.stepfun.modelling_step37.transformer_block import get_step37_text_layer_spec
 from megatron.bridge.models.stepfun.step35_bridge import build_step35_layer_spec
 from megatron.bridge.models.stepfun.step37_bridge import Step37Bridge
@@ -96,6 +99,27 @@ def test_checkpoint_mtp_override_is_nested_in_exported_config(self):
 
         assert synthesized_config.text_config.num_nextn_predict_layers == 0
 
+    def test_projector_bias_schema_roundtrips(self):
+        reference_config = Step37Config(projector_bias=False)
+        provider = SimpleNamespace(projector_bias=True)
+
+        biased_projector = ImageInsertEmbedding(nn.Identity(), 4, 2, projector_bias=provider.projector_bias)
+        unbiased_projector = ImageInsertEmbedding(nn.Identity(), 4, 2, projector_bias=False)
+        assert biased_projector.align_projector.bias is not None
+        assert unbiased_projector.align_projector.bias is None
+
+        mapping = (
+            Step37Bridge().mapping_registry().megatron_to_hf_lookup("image_insert_embedding.align_projector.bias")
+        )
+        checkpoint_config = Step37Bridge.megatron_to_hf_config(provider)
+        exported_config = conform_config_to_reference(checkpoint_config, reference_config.to_dict())
+        synthesized_config = Step37Config(**exported_config)
+
+        assert (
+            mapping.hf_param if mapping is not None else None,
+            synthesized_config.projector_bias,
+        ) == ("vit_large_projector.bias", True)
+
 
 class TestGetStep37TextLayerSpec:
     def test_delegates_to_step35_builder(self):