Update modeling_neollm.py

modeling_neollm.py

@@ -4562,7 +4562,7 @@ class NeoLLMDecoderLayer(GradientCheckpointingLayer):
         self.register_buffer(
             "_siamese_stream_scale_value",
             torch.tensor(siamese_stream_scale, dtype=torch.float32),
-            persistent=False,
+            persistent=True,
         )
         # Controls only the first pre-attention normalisation applied directly
         # to the embedding stream. Defaults to True for checkpoint/config