diff --git a/python/sglang/srt/managers/schedule_batch.py b/python/sglang/srt/managers/schedule_batch.py
index 12afdfd74..23d6dfc35 100644
--- a/python/sglang/srt/managers/schedule_batch.py
+++ b/python/sglang/srt/managers/schedule_batch.py
@@ -362,6 +362,7 @@ class MultimodalInputs:
     # QWen2-VL related
     mrope_positions: Optional[torch.Tensor] = None
     mrope_position_delta: Optional[torch.Tensor] = None
+    mrope_position_delta_repeated_cache: Optional[torch.Tensor] = None
 
     @staticmethod
     def from_dict(obj: dict):
diff --git a/python/sglang/srt/model_executor/forward_batch_info.py b/python/sglang/srt/model_executor/forward_batch_info.py
index a713f399b..aaeb4fa7f 100644
--- a/python/sglang/srt/model_executor/forward_batch_info.py
+++ b/python/sglang/srt/model_executor/forward_batch_info.py
@@ -654,10 +654,11 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
         seq_len: int,
     ) -> torch.Tensor:
         # doing below compute on cpu to avoid frequent small kernels
-        mrope_position_deltas = mm_input.mrope_position_delta.flatten()
-        mrope_positions = (
-            (mrope_position_deltas + seq_len - 1).unsqueeze(0).repeat(3, 1)
-        )
+        if mm_input.mrope_position_delta_repeated_cache is None:
+            mm_input.mrope_position_delta_repeated_cache = (
+                (mm_input.mrope_position_delta - 1).flatten().unsqueeze(0).repeat(3, 1)
+            )
+        mrope_positions = mm_input.mrope_position_delta_repeated_cache + seq_len
         return mrope_positions
 
     def _compute_mrope_positions(