diff --git a/python/sglang/srt/disaggregation/decode.py b/python/sglang/srt/disaggregation/decode.py
index c55ff312b..f5be51e83 100644
--- a/python/sglang/srt/disaggregation/decode.py
+++ b/python/sglang/srt/disaggregation/decode.py
@@ -421,6 +421,7 @@ class DecodePreallocQueue:
 
             # Fast path: cache-only lookup, no network calls
             prefill_dp_rank = self._resolve_prefill_dp_rank(req)
+            logger.debug(f"prefill_dp_rank: {prefill_dp_rank}")
             if prefill_dp_rank is not None:
                 decode_req.kv_receiver.init(prefill_dp_rank)
                 return
@@ -428,13 +429,14 @@ class DecodePreallocQueue:
             self.pending_reqs.append(decode_req)
 
     def _resolve_prefill_dp_rank(self, req: Req) -> Optional[int]:
-        if req.disagg_prefill_dp_rank is not None:
-            return req.disagg_prefill_dp_rank
-
         prefill_info = self.kv_manager.prefill_info_table.get(_bootstrap_addr(req))
+        # If None, it will go to the slow path and resolve prefill_info by _ensure_prefill_info then cache it
         if prefill_info is None:
             return None
 
+        if req.disagg_prefill_dp_rank is not None:
+            return req.disagg_prefill_dp_rank
+
         if prefill_info.dp_size == 1:
             return 0
 
diff --git a/test/registered/unit/managers/test_hisparse_unit.py b/test/registered/unit/managers/test_hisparse_unit.py
index 8456d91ed..6c4028be6 100644
--- a/test/registered/unit/managers/test_hisparse_unit.py
+++ b/test/registered/unit/managers/test_hisparse_unit.py
@@ -305,10 +305,14 @@ class TestHiSparseUnit(unittest.TestCase):
             for i in range(TOP_K):
                 if batch[b, i] < 0:
                     continue
-                nd = self.device_pool.kv_buffer[layer_id][naive_locs[b, i].long()]
-                kd = self.device_pool.kv_buffer[layer_id][kernel_locs[b, i].long()]
+                naive_data = self.device_pool.kv_buffer[layer_id][
+                    naive_locs[b, i].long()
+                ]
+                kernel_data = self.device_pool.kv_buffer[layer_id][
+                    kernel_locs[b, i].long()
+                ]
                 self.assertTrue(
-                    torch.allclose(nd.float(), kd.float(), atol=1e-2),
+                    torch.allclose(naive_data.float(), kernel_data.float(), atol=1e-2),
                     f"{msg}layer {layer_id}, b{b} idx {i}: naive != kernel",
                 )