Fix gemma3 monkey patch tests (#735)

## Summary
With transformers >= 4.52.0, there was some refactoring of gemma3 model
code. Patching should still work for previous transformers version, but
gemma3 tests won't pass for older transformers versions. Not sure we
want to maintain that logic in the tests.

For reference:

- Before (transformers <=4.51.3):
```
Gemma3ForConditionalGeneration
 - language_model (Gemma3ForCausalLM)
    - model (Gemma3TextModel)
       - layers/norm/etc.
```

- After:
```
Gemma3ForConditionalGeneration
- model (Gemma3Model)
   - language_model (Gemma3TextModel)
      - layers/norm/etc.
- language_model (for backwards-compatibility, points to model.language_model (Gemma3TextModel))
```

Fixing part of #729

## Testing Done
Gemma3 monkey patch tests pass
```
pytest test/transformers/test_monkey_patch.py::test_apply_liger_kernel_to_instance_for_gemma3_conditional_generation
pytest test/transformers/test_monkey_patch.py::test_apply_liger_kernel_to_instance_for_gemma3_text 
```

- Hardware Type: <BLANK>
- [ ] run `make test` to ensure correctness
- [ ] run `make checkstyle` to ensure code style
- [ ] run `make test-convergence` to ensure convergence

Author: shimizust

src/liger_kernel/transformers/monkey_patch.py

@@ -776,7 +776,7 @@ def apply_liger_kernel_to_gemma3_text(
 
     from transformers.models.gemma3 import modeling_gemma3
     from transformers.models.gemma3.modeling_gemma3 import Gemma3DecoderLayer
-    from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM
+    from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM, Gemma3TextModel
 
     from liger_kernel.transformers.gema3_rms import LigerRMSNormForGemma3
     from liger_kernel.transformers.model.gemma3 import causal_forward
@@ -807,9 +807,9 @@ def apply_liger_kernel_to_gemma3_text(
         # The model instance already exists, so we need to additionally patch the
         # instance variables that reference already-instantiated modules
 
-        if isinstance(model, Gemma3ForCausalLM):
+        if isinstance(model, Gemma3ForCausalLM) or isinstance(model, Gemma3TextModel):
             # get the base model from the model instance
-            base_model = model.model
+            base_model = model.model if isinstance(model, Gemma3ForCausalLM) else model
 
             if rms_norm:
                 _patch_rms_norm_module_for_gemma3(base_model.norm)
@@ -1625,7 +1625,6 @@ def _apply_liger_kernel_to_instance(model: PreTrainedModel, **kwargs) -> None:
         return
 
     apply_fn = MODEL_TYPE_TO_APPLY_LIGER_FN[model_type]
-
     apply_fn_signature = inspect.signature(apply_fn)
 
     # Filter out the keyword arguments that are not supported by the apply function

test/transformers/test_monkey_patch.py

@@ -667,7 +667,7 @@ def test_apply_liger_kernel_to_instance_for_gemma3_text():
 
 
 @pytest.mark.skipif(not is_gemma3_available(), reason="gemma3 module not available")
-def test_apply_liger_kernel_to_instance_for_gemma3():
+def test_apply_liger_kernel_to_instance_for_gemma3_conditional_generation():
     # Ensure any monkey patching is cleaned up for subsequent tests
 
     with patch("transformers.models.gemma3.modeling_gemma3"):
@@ -687,8 +687,8 @@ def test_apply_liger_kernel_to_instance_for_gemma3():
             intermediate_size=64,
         )
         config = transformers.models.gemma3.configuration_gemma3.Gemma3Config(text_config, vision_config)
-        dummy_model_instance = Gemma3ForConditionalGeneration._from_config(config)
 
+        dummy_model_instance = Gemma3ForConditionalGeneration._from_config(config)
         assert isinstance(dummy_model_instance, Gemma3ForConditionalGeneration)
 
         # Check that model instance variables are not yet patched with Liger modules
@@ -704,11 +704,11 @@ def test_apply_liger_kernel_to_instance_for_gemma3():
             dummy_model_instance.multi_modal_projector.mm_soft_emb_norm.forward
         ) != inspect.getsource(LigerRMSNorm.forward)
 
-        assert inspect.getsource(dummy_model_instance.language_model.model.norm.forward) != inspect.getsource(
+        assert inspect.getsource(dummy_model_instance.language_model.norm.forward) != inspect.getsource(
             LigerRMSNorm.forward
         )
 
-        for layer in dummy_model_instance.language_model.model.layers:
+        for layer in dummy_model_instance.language_model.layers:
             assert inspect.getsource(layer.mlp.forward) != inspect.getsource(LigerGEGLUMLP.forward)
             assert inspect.getsource(layer.input_layernorm.forward) != inspect.getsource(LigerRMSNorm.forward)
             assert inspect.getsource(layer.post_attention_layernorm.forward) != inspect.getsource(LigerRMSNorm.forward)
@@ -736,10 +736,10 @@ def test_apply_liger_kernel_to_instance_for_gemma3():
             dummy_model_instance.multi_modal_projector.mm_soft_emb_norm.forward
         ) == inspect.getsource(LigerRMSNorm.forward)
 
-        assert inspect.getsource(dummy_model_instance.language_model.model.norm.forward) == inspect.getsource(
+        assert inspect.getsource(dummy_model_instance.language_model.norm.forward) == inspect.getsource(
             LigerRMSNorm.forward
         )
-        for layer in dummy_model_instance.language_model.model.layers:
+        for layer in dummy_model_instance.language_model.layers:
             assert inspect.getsource(layer.mlp.forward) == inspect.getsource(LigerGEGLUMLP.forward)
             assert inspect.getsource(layer.input_layernorm.forward) == inspect.getsource(LigerRMSNorm.forward)
             assert inspect.getsource(layer.post_attention_layernorm.forward) == inspect.getsource(LigerRMSNorm.forward)