Fix Qwen2.5-VL huggingface conversion issue (#2107)

src/megatron/bridge/inference/vlm/base.py

@@ -112,6 +112,17 @@ def setup_model_and_tokenizer(
     return inference_wrapped_model, processor
 
 
+def _expose_decoder_from_language_model(model):
+    """Recursively get language_model from model and expose decoder, handling wrapped modules."""
+    current = model
+    while hasattr(current, "module"):
+        current = current.module
+
+    if hasattr(current, "language_model"):
+        language_model = current.language_model
+        current.decoder = language_model.decoder
+
+
 def setup_inference_wrapper(
     model,
     tokenizer,
@@ -131,6 +142,8 @@ def setup_inference_wrapper(
         wrapper_cls = QwenVLInferenceWrapper
         if isinstance(config, Qwen25VLModelProvider):
             hidden_size = config.hidden_size
+            # Expose decoder for MCore Infernce Engine compatibility (used by get_mamba_inference_state_config_from_model)
+            _expose_decoder_from_language_model(mcore_model)
         else:
             hidden_size = config.language_transformer_config.hidden_size
     else:

src/megatron/bridge/models/qwen_vl/modeling_qwen25_vl.py

@@ -111,9 +111,6 @@ def __init__(
         self.share_embeddings_and_output_weights = config.share_embeddings_and_output_weights
         self.shared_embedding_or_output_weight = self.language_model.shared_embedding_or_output_weight
 
-        # Expose decoder for MCore Infernce Engine compatibility (used by get_mamba_inference_state_config_from_model)
-        self.decoder = self.language_model.decoder
-
         # Bind methods from HF's Qwen2_5_VLModel to this instance
         # get_placeholder_mask is only available in transformers 4.55+
         if is_transformers_min_version("4.55.0"):

tests/unit_tests/inference/vlm/test_base.py

@@ -236,25 +236,54 @@ class TestSetupInferenceWrapper:
 
     @patch("megatron.bridge.inference.vlm.base.QwenVLInferenceWrapper")
     def test_setup_inference_wrapper_qwen25(self, mock_wrapper_cls, mock_tokenizer):
-        mock_model = MagicMock()
+        """Test Qwen25 setup with module.language_model.decoder structure."""
+
+        # Create mock objects with nested structure
+        class MockObject:
+            pass
+
+        mock_decoder = MagicMock()
+
+        # Build the nested structure: model.module.language_model.decoder
+        mock_language_model = MockObject()
+        mock_language_model.decoder = mock_decoder
+
+        mock_module = MockObject()
+        mock_module.language_model = mock_language_model
+
+        mock_model = MockObject()
+        mock_model.module = mock_module
         mock_model.config = MagicMock(spec=Qwen25VLModelProvider)
         mock_model.config.hidden_size = 1024
+        mock_model.cuda = MagicMock(return_value=mock_model)
+        mock_model.to = MagicMock(return_value=mock_model)
+        mock_model.eval = MagicMock()
 
         _wrapper = setup_inference_wrapper(mock_model, mock_tokenizer)
 
+        # Verify decoder was exposed at module level
+        assert hasattr(mock_module, "decoder")
+        assert mock_module.decoder is mock_decoder
+
         mock_wrapper_cls.assert_called_once()
         # Check InferenceWrapperConfig was created with correct hidden_size
-        # Args are positional: (model, InferenceWrapperConfig)
         call_args = mock_wrapper_cls.call_args
-        inference_config = call_args[0][1]  # Second positional argument
+        inference_config = call_args[0][1]
         assert inference_config.hidden_size == 1024
 
     @patch("megatron.bridge.inference.vlm.base.QwenVLInferenceWrapper")
     def test_setup_inference_wrapper_qwen3(self, mock_wrapper_cls, mock_tokenizer):
-        mock_model = MagicMock()
+        # Create a simple object without module attribute to avoid infinite loop
+        class MockObject:
+            pass
+
+        mock_model = MockObject()
         mock_model.config = MagicMock(spec=Qwen3VLModelProvider)
         mock_model.config.language_transformer_config = MagicMock()
         mock_model.config.language_transformer_config.hidden_size = 2048
+        mock_model.cuda = MagicMock(return_value=mock_model)
+        mock_model.to = MagicMock(return_value=mock_model)
+        mock_model.eval = MagicMock()
 
         _wrapper = setup_inference_wrapper(mock_model, mock_tokenizer)
 
@@ -266,8 +295,15 @@ def test_setup_inference_wrapper_qwen3(self, mock_wrapper_cls, mock_tokenizer):
         assert inference_config.hidden_size == 2048
 
     def test_setup_inference_wrapper_invalid(self, mock_tokenizer):
-        mock_model = MagicMock()
+        # Create a simple object without module attribute to avoid infinite loop
+        class MockObject:
+            pass
+
+        mock_model = MockObject()
         mock_model.config = MagicMock()  # Not Qwen config
+        mock_model.cuda = MagicMock(return_value=mock_model)
+        mock_model.to = MagicMock(return_value=mock_model)
+        mock_model.eval = MagicMock()
 
         with pytest.raises(ValueError):
             setup_inference_wrapper(mock_model, mock_tokenizer)