Update model.py

model.py

@@ -355,6 +355,8 @@ class VoilaModel(LlamaPreTrainedModel):
         if self.training or \
                 (past_key_values is None and ref_embs is not None) or \
                 (past_key_values is not None and past_key_values.get_seq_length() < 4 and ref_embs is not None):
+            print("model device",self.ref_emb_linear.weight.device)
+            print('tensor device', ref_embs.device)
             ref_embs = self.ref_emb_linear(ref_embs.to(self.ref_emb_linear.weight.dtype))
             ref_embs = ref_embs * ref_embs_mask.unsqueeze(-1).unsqueeze(-1)
             # (padding_left,padding_right,padding_top,padding_bottom,padding_front,padding_back)