diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.data2vec_audio.py.swp b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.data2vec_audio.py.swp
index 0b648eee8..9706d9a0c 100644
Binary files a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.data2vec_audio.py.swp and b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.data2vec_audio.py.swp differ
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/data2vec_audio.py b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/data2vec_audio.py
index 5b8d4dedf..3dcac229c 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/data2vec_audio.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/data2vec_audio.py
@@ -40,7 +40,7 @@ class TransformerEncoderAdapter(TransformerEncoder):
         super().__init__(args)
         self.adapters = ResidualAdapterModule()
         for p in self.adapters.parameters():
-            p.data = torch.nn.parameters(0.)
+            p.data = torch.nn.parameters(torch.zeros(p.size()).to('cuda'))
 
     def forward(self, x, padding_mask=None, layer=None, tgt_layer=None):
         x, layer_results = self.extract_features_with_adapter(