Update inferencer.py

inferencer.py

@@ -186,6 +186,13 @@ class UniPicV2Inferencer:
         attention_mask = torch.cat([attention_mask, pad_mask], dim=1)
         
         # Get input embeddings
+        
+        # 获取 embedding 权重所在设备
+        embed_device = self.lmm.get_input_embeddings().weight.device
+        
+        # 确保 input_ids 在同一设备
+        input_ids = input_ids.to(embed_device)
+        
         inputs_embeds = self.lmm.get_input_embeddings()(input_ids)
         
         # Ensure meta queries are on correct device