[Unified Tree] Preserve aux LRU recency when splitting nodes (#38482)

python/sglang/srt/mem_cache/unified_cache/components/swa_component.py

@@ -615,7 +615,10 @@ class SWAComponent(TreeComponent):
                 new_parent.component_data[self.component_type].value is None
                 and parent_swa_data.host_lock_ref == 0
             ):
-                host_lru.insert_mru(new_parent)
+                if host_lru.in_list(child):
+                    host_lru.insert_after(child, new_parent)
+                else:
+                    host_lru.insert_mru(new_parent)
             if (
                 child.component_data[self.component_type].value is None
                 and child_swa_data.host_lock_ref == 0

python/sglang/srt/mem_cache/unified_cache/unified_tree_core.py

@@ -239,6 +239,12 @@ class UnifiedLRUList:
         self.cache[node.id] = node
         self._add_node(node)
 
+    def insert_after(self, prev_node: UnifiedTreeNode, node: UnifiedTreeNode):
+        assert prev_node.id in self.cache
+        assert node.id not in self.cache
+        self.cache[node.id] = node
+        self._add_node_after(prev_node, node)
+
     def remove_node(self, node: UnifiedTreeNode):
         assert node.id in self.cache
         del self.cache[node.id]
@@ -1307,8 +1313,6 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
         # owner (b + P) % N on both sides of the split).
         new_node.rotation_base = child.rotation_base
 
-        self._for_each_component_lru(child, UnifiedLRUList.remove_node)
-
         child.parent = new_node
         child.key = child.key[split_len:]
         new_node.hash_value, child.hash_value = split_node_hash_value(
@@ -1334,11 +1338,12 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
                 new_child_node_id=child.id,
             )
 
+        # Splitting does not access the suffix; retain its recency and place
+        # the inherited prefix beside it, in the same session partition.
         self._for_each_component_lru(
-            new_node, UnifiedLRUList.insert_mru, skip_existing=True
-        )
-        self._for_each_component_lru(
-            child, UnifiedLRUList.insert_mru, skip_existing=True
+            new_node,
+            lambda lru, node: lru.insert_after(child, node),
+            skip_existing=True,
         )
         child.last_access_time = get_and_increase_time_counter()