Bugfix for minicpmo vision test (#5760)

python/sglang/srt/managers/multimodal_processors/minicpm.py

@@ -97,7 +97,7 @@ class MiniCPMMultimodalProcessor(BaseMultimodalProcessor):
             audio_start_id = tokenizer.audio_start_id
             audio_end_id = tokenizer.audio_end_id
 
-        im_token_id = tokenizer.unk_token_id
+        im_token_id = tokenizer.unk_id
         pixel_values = res["pixel_values"]
         tgt_sizes = res["tgt_sizes"]
 

python/sglang/srt/managers/schedule_batch.py

@@ -374,6 +374,12 @@ class MultimodalInputs:
                 self.mrope_position_delta = torch.cat(
                     [self.mrope_position_delta, other.mrope_position_delta], dim=0
                 )
+
+        for key, val in other.__dict__.items():
+            if "_id" in key:
+                # set token_ids
+                if getattr(self, key, None) is None:
+                    setattr(self, key, getattr(other, key, None))
         # other args would be kept intact