[MM] MultimodalInputs.merge: adopt the M-RoPE delta and clear its decode cache

python/sglang/srt/managers/schedule_batch.py

@@ -888,6 +888,9 @@ def merge(self, other: MultimodalInputs):
                 self.mrope_position_delta = torch.cat(
                     [self.mrope_position_delta, other.mrope_position_delta], dim=0
                 )
+        elif other.mrope_position_delta is not None:
+            self.mrope_position_delta = other.mrope_position_delta
+        self.mrope_position_delta_repeated_cache = None
 
         for key, val in other.__dict__.items():
             if "_id" in key:

test/registered/unit/managers/test_mm_inputs_merge_mrope_delta.py

@@ -0,0 +1,43 @@
+import unittest
+
+import torch
+
+from sglang.srt.managers.schedule_batch import (
+    Modality,
+    MultimodalDataItem,
+    MultimodalInputs,
+)
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+IM_TOKEN_ID = 7
+
+
+class TestMultimodalInputsMergeMropeDelta(CustomTestCase):
+    def test_merge_invalidates_mrope_delta_cache(self):
+        def _mm(delta):
+            item = MultimodalDataItem(
+                modality=Modality.IMAGE, hash=7, pad_value=7, offsets=[(0, 0)]
+            )
+            return MultimodalInputs(
+                mm_items=[item],
+                im_token_id=IM_TOKEN_ID,
+                mrope_position_delta=torch.tensor([[delta]]),
+            )
+
+        base = _mm(3)
+        base.mrope_position_delta_repeated_cache = torch.zeros(3, 1, dtype=torch.long)
+        base.merge(_mm(5))
+        self.assertIsNone(base.mrope_position_delta_repeated_cache)
+        self.assertEqual(base.mrope_position_delta.flatten().tolist(), [3, 5])
+
+        no_delta = _mm(1)
+        no_delta.mrope_position_delta = None
+        no_delta.merge(_mm(9))
+        self.assertEqual(no_delta.mrope_position_delta.flatten().tolist(), [9])
+
+
+if __name__ == "__main__":
+    unittest.main()