[https://nvbugs/5429689][fix] Fix mllama model structure update with transformers issue (#6699)

Signed-off-by: Wangshanshan <[email protected]>

Author: dominicshanshan

tensorrt_llm/tools/multimodal_builder.py

@@ -1188,8 +1188,18 @@ def forward(self, pixel_values, aspect_ratio_ids, aspect_ratio_mask):
     model = MllamaForConditionalGeneration.from_pretrained(args.model_path,
                                                            torch_dtype='auto',
                                                            device_map='auto')
-    wrapper = MLLaMAVisionWrapper(model.vision_model,
-                                  model.multi_modal_projector)
+
+    # Check if the model structure is updated to transformers >= 4.52.0
+    if hasattr(model, 'model') and hasattr(model.model, 'vision_model'):
+        vision_model = model.model.vision_model
+        multi_modal_projector = model.model.multi_modal_projector
+    else:
+        # transformers < 4.52.0
+        vision_model = model.vision_model
+        multi_modal_projector = model.multi_modal_projector
+
+    wrapper = MLLaMAVisionWrapper(vision_model, multi_modal_projector)
+
     model_dtype = model.dtype
     image = Image.new('RGB', [2048, 2688])  # dummy image
     inputs = processor(images=image,