Merge branch 'main' into fix-gkd-liger-mem-spike

Author: kashif

tests/test_gkd_trainer.py

@@ -14,6 +14,7 @@
 
 import os
 
+import pytest
 import torch
 import torch.nn.functional as F
 from datasets import load_dataset
@@ -246,6 +247,7 @@ def test_gkd_trainer(self):
         self.assertIn("model.safetensors", os.listdir(self.tmp_dir + "/checkpoint-2"))
 
     @require_liger_kernel
+    @pytest.mark.xfail(reason="Computing the Liger loss spikes GPU memory usage, causing the test to run OOM.")
     def test_gkd_trainer_with_liger(self):
         training_args = GKDConfig(
             output_dir=self.tmp_dir,