diff --git a/tests/tp/test_tp.py b/tests/tp/test_tp.py
index bb431cef9..586de4f5b 100644
--- a/tests/tp/test_tp.py
+++ b/tests/tp/test_tp.py
@@ -84,8 +84,8 @@ class TestTensorParallel(TestCasePlus):
             expected_model_memory = 16
             overhead_factor = 1.2
 
-            # # Check that we do not use more than the expected sharded size during initialization
-            if not torch.cuda.max_memory_allocated(device) / 1024**3 < (expected_model_memory / world_size) * overhead_factor:
+            # Check that we do not use more than the expected sharded size during initialization
+            if torch.cuda.max_memory_allocated(device) / 1024**3 > (expected_model_memory / world_size) * overhead_factor:
                 raise ValueError("Loading the model used more than the expected fraction of model size per device")
 
             torch.distributed.barrier()