fix CI issue for vlm_gemma_3n model (#4278)

Signed-off-by: Liu, Kaixuan <[email protected]>
Co-authored-by: Quentin Gallouédec <[email protected]>
Co-authored-by: Albert Villanova del Moral <[email protected]>
Co-authored-by: Quentin Gallouédec <[email protected]>

Author: 3 people

tests/test_sft_trainer.py

@@ -1494,6 +1494,7 @@ def test_train_vlm_gemma_3n(self):
         # Initialize the trainer
         training_args = SFTConfig(
             output_dir=self.tmp_dir,
+            learning_rate=0.1,
             max_length=None,
             per_device_train_batch_size=1,
             gradient_checkpointing=True,
@@ -1514,8 +1515,8 @@ def test_train_vlm_gemma_3n(self):
         # Check the params have changed
         for n, param in previous_trainable_params.items():
             new_param = trainer.model.get_parameter(n)
-            if "model.vision_tower" in n:
-                # The vision tower is not updated, not sure why at this point.
+            if "model.audio_tower" in n or "model.embed_audio" in n:
+                # The audio embedding parameters are not updated because this dataset contains no audio data
                 continue
             assert not torch.allclose(param, new_param, rtol=1e-12, atol=1e-12), f"Param {n} is not updated"