Update infer.py

infer.py

@@ -63,6 +63,8 @@ def eval_model(model, tokenizer, tokenizer_voila, model_type, task_type, history
     # step1: initializing
     model.to('cuda')
     tokenizer_voila.to('cuda')
+    ref_embs = ref_embs.to('cuda')
+    ref_embs_mask = ref_embs_mask.to('cuda')
     num_codebooks = model.config.num_codebooks
     codebook_size = model.config.codebook_size