diff --git a/tests/test_shadow_peft.py b/tests/test_shadow_peft.py
index ffae772..7ae5ece 100644
--- a/tests/test_shadow_peft.py
+++ b/tests/test_shadow_peft.py
@@ -87,6 +87,7 @@ def test_save_load_roundtrip_matches_outputs(tmp_path: Path):
     torch.manual_seed(0)
     base1 = _tiny_llama(num_layers=4)
     base1.eval()
+    base1_state = {k: v.detach().clone() for k, v in base1.state_dict().items()}
 
     cfg = ShadowConfig(
         num_shadow_layers=1,
@@ -107,7 +108,7 @@ def test_save_load_roundtrip_matches_outputs(tmp_path: Path):
 
     # Create a new base with identical weights.
     base2 = _tiny_llama(num_layers=4)
-    base2.load_state_dict(base1.state_dict())
+    base2.load_state_dict(base1_state)
     base2.eval()
 
     m2 = ShadowPeftModel.from_pretrained(base2, save_dir, is_trainable=False)
@@ -217,5 +218,3 @@ def test_implicit_shadow_model_uses_shadow_intermediate_size():
     )
     m = get_shadow_model(base, cfg)
     assert m.shadow_model.config.intermediate_size == 12
-
-