requirements

WavLM_modules.py

@@ -450,8 +450,7 @@ class MultiheadAttention(nn.Module):
         relative_position_bucket = self._relative_positions_bucket(
             relative_position, bidirectional=True
         )
-        relative_position_bucket.cuda()
-        values = self.relative_attention_bias(relative_position_bucket)
+        values = self.relative_attention_bias(relative_position_bucket.cuda())
         values = values.permute([2, 0, 1])
         return values