From ba1f1dc132257f0e92d2266f966dc19758a3f2b3 Mon Sep 17 00:00:00 2001
From: Steven Shimizu <shimizust@gmail.com>
Date: Mon, 16 Sep 2024 17:40:24 -0700
Subject: [PATCH] Updated Trainer's liger-kernel integration to call correct
 patching API (#33502)

* Updated liger-kernel integration in Trainer to call correct patching API

* Fixed styling
---
 src/transformers/trainer.py            | 13 ++++++-----
 src/transformers/utils/import_utils.py |  2 +-
 tests/trainer/test_trainer.py          | 30 +++++++++++++++-----------
 3 files changed, 25 insertions(+), 20 deletions(-)

diff --git a/src/transformers/trainer.py b/src/transformers/trainer.py
index f815c50d59..97a0520936 100755
--- a/src/transformers/trainer.py
+++ b/src/transformers/trainer.py
@@ -468,19 +468,18 @@ class Trainer:
 
         if self.args.use_liger_kernel:
             if is_liger_kernel_available():
-                from liger_kernel.transformers.trainer_integration import _apply_liger_kernel
+                from liger_kernel.transformers import _apply_liger_kernel_to_instance
 
-                model_type = getattr(model, "config", None) and getattr(model.config, "model_type", None)
-                if model_type:
-                    # Monkey patch the model with liger kernels. Use the default kernel configurations.
-                    _apply_liger_kernel(model_type=model_type)
+                if isinstance(model, PreTrainedModel):
+                    # Patch the model with liger kernels. Use the default kernel configurations.
+                    _apply_liger_kernel_to_instance(model=model)
                 else:
                     logger.warning(
-                        "The model does not have a valid `model_type` specified. No liger kernels will be applied."
+                        "The model is not an instance of PreTrainedModel. No liger kernels will be applied."
                     )
             else:
                 raise ImportError(
-                    "You have set `use_liger_kernel` to `True` but liger-kernel >= 0.1.0 is not available. "
+                    "You have set `use_liger_kernel` to `True` but liger-kernel >= 0.3.0 is not available. "
                     "Please install it with `pip install liger-kernel`"
                 )
 
diff --git a/src/transformers/utils/import_utils.py b/src/transformers/utils/import_utils.py
index af8a569443..ad8b649aaa 100755
--- a/src/transformers/utils/import_utils.py
+++ b/src/transformers/utils/import_utils.py
@@ -1187,7 +1187,7 @@ def is_liger_kernel_available():
     if not _liger_kernel_available:
         return False
 
-    return version.parse(importlib.metadata.version("liger_kernel")) >= version.parse("0.1.0")
+    return version.parse(importlib.metadata.version("liger_kernel")) >= version.parse("0.3.0")
 
 
 # docstyle-ignore
diff --git a/tests/trainer/test_trainer.py b/tests/trainer/test_trainer.py
index 1837d98903..791486ec83 100644
--- a/tests/trainer/test_trainer.py
+++ b/tests/trainer/test_trainer.py
@@ -1344,22 +1344,28 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
 
     @require_liger_kernel
     def test_use_liger_kernel_patching(self):
-        # Test that the model code actually gets patched with Liger kernel
-        from liger_kernel.transformers.rms_norm import LigerRMSNorm
+        # Ensure any monkey patching is cleaned up for subsequent tests
+        with patch("transformers.models.llama.modeling_llama"):
+            from liger_kernel.transformers import LigerRMSNorm, liger_rotary_pos_emb
 
-        from transformers.models.llama import modeling_llama
+            from transformers.models.llama import modeling_llama
 
-        config = LlamaConfig(vocab_size=100, hidden_size=32, num_hidden_layers=3, num_attention_heads=4)
-        tiny_llama = LlamaForCausalLM(config)
+            config = LlamaConfig(vocab_size=100, hidden_size=32, num_hidden_layers=3, num_attention_heads=4)
+            tiny_llama = LlamaForCausalLM(config)
 
-        args = TrainingArguments(
-            "./test",
-            use_liger_kernel=True,
-        )
-        Trainer(tiny_llama, args)
+            # Spot check that modeling code and model instance variables are not yet patched
+            self.assertNotEqual(modeling_llama.apply_rotary_pos_emb, liger_rotary_pos_emb)
+            self.assertFalse(isinstance(tiny_llama.model.norm, LigerRMSNorm))
 
-        # Check that one of the Llama model layers has been correctly patched with Liger kernel
-        self.assertEqual(modeling_llama.LlamaRMSNorm, LigerRMSNorm)
+            args = TrainingArguments(
+                "./test",
+                use_liger_kernel=True,
+            )
+            Trainer(tiny_llama, args)
+
+            # Spot check that modeling code and model instance variables are patched
+            self.assertEqual(modeling_llama.apply_rotary_pos_emb, liger_rotary_pos_emb)
+            self.assertTrue(isinstance(tiny_llama.model.norm, LigerRMSNorm))
 
     @require_liger_kernel
     @require_torch_gpu