Update modeling_diva.py

modeling_diva.py

@@ -284,7 +284,9 @@ class DiVAModel(PreTrainedModel):
         while greedy != 128009 and len(outs) < max_new_tokens:
             past_key_values = outputs.past_key_values if outputs else None
             outputs = self.llama_decoder(
-                inputs_embeds=inputs_embeds,
+                inputs_embeds=.to(
+                    self.llama_decoder.model.embed_tokens.weight.device
+                ).half(),
                 return_dict=True,
                 output_hidden_states=True,
                 past_key_values=past_key_values,