latest fixes needed for GA in latest transformers

tests/e2e/patched/test_mixtral_samplepack.py

@@ -102,9 +102,5 @@ class TestMixtral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        model, _ = train(cfg=cfg, dataset_meta=dataset_meta)
-        assert (
-            "MixtralFlashAttention2"
-            in model.model.layers[0].self_attn.__class__.__name__
-        )
+        train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_model_patches.py

@@ -49,12 +49,7 @@ class TestModelPatches(unittest.TestCase):
         )
         normalize_config(cfg)
         tokenizer = load_tokenizer(cfg)
-        model, _ = load_model(cfg, tokenizer, inference=False)
-
-        assert (
-            "MixtralFlashAttention2"
-            in model.model.layers[0].self_attn.__class__.__name__
-        )
+        load_model(cfg, tokenizer, inference=False)
 
     @with_temp_dir
     def test_mistral_multipack(self, temp_dir):

tests/patched/test_llama_trainer_ga.py

@@ -9,7 +9,6 @@ from axolotl.monkeypatch.trainer_grad_accum import (
 )
 
 
-@pytest.mark.skip("may not be needed for latest transformers version")
 class TestTrainerGAIntegration(unittest.TestCase):
     """llama monkeypatch integration tests."""
 
@@ -20,6 +19,7 @@ class TestTrainerGAIntegration(unittest.TestCase):
             "HF transformers Trainer.training_step has changed and isn't patchable",
         )
 
+    @pytest.mark.skip("may not be needed for latest transformers version")
     def test_model_forward_patchable(self):
         # ensures the current version of transformers has loss code that matches our patching code
         self.assertTrue(