Catch the cross-architecture warning in the GOLD teacher VLM inputs test (#7728)

Author: albertvillanova

tests/experimental/test_gold_trainer.py

@@ -1576,12 +1576,13 @@ def test_build_teacher_vlm_inputs_feeds_images_and_completion_byte_offsets(tmp_p
     """Cross-architecture VLM ULD must render the teacher prompt through the teacher processor (so it
     actually sees the image via pixel_values) while keeping completion byte offsets relative to the original completion
     text — the coordinate system shared with the student."""
-    trainer = GOLDTrainer(
-        model=_TINY_SMOLVLM,
-        teacher_model=_TINY_QWEN3_VL,
-        args=GOLDConfig(output_dir=str(tmp_path), report_to="none", use_uld_loss=True, use_extended_uld=True),
-        train_dataset=vlm_dataset,
-    )
+    with pytest.warns(UserWarning, match="Cross-architecture VLM distillation"):
+        trainer = GOLDTrainer(
+            model=_TINY_SMOLVLM,
+            teacher_model=_TINY_QWEN3_VL,
+            args=GOLDConfig(output_dir=str(tmp_path), report_to="none", use_uld_loss=True, use_extended_uld=True),
+            train_dataset=vlm_dataset,
+        )
 
     images, prompts = trainer._extract_images_and_prompts(vlm_examples)
     completion_texts = _get_assistant_texts(vlm_examples)