[BUGFIX] fix value oom in radix tree (sgl-project#17400)

Author: yizhang2077

python/sglang/srt/mem_cache/hiradix_cache.py

@@ -796,7 +796,7 @@ def _insert_helper_host(
             new_node.parent = node
             new_node.key = key
             new_node.value = None
-            new_node.host_value = host_value
+            new_node.host_value = host_value.clone()
             new_node.hash_value = hash_value
             node.children[child_key] = new_node
         return matched_length
@@ -897,7 +897,7 @@ def insert(
                 # shared-prefix node should also reflect max priority
                 new_node.priority = max(new_node.priority, priority)
                 if new_node.evicted:
-                    new_node.value = value[:prefix_len]
+                    new_node.value = value[:prefix_len].clone()
                     self.evictable_size_ += len(new_node.value)
                 else:
                     self._inc_hit_count(new_node, chunked)
@@ -914,7 +914,7 @@ def insert(
             new_node = TreeNode(priority=priority)
             new_node.parent = node
             new_node.key = key
-            new_node.value = value
+            new_node.value = value.clone()
             node.children[child_key] = new_node
             self.evictable_size_ += len(value)
 

python/sglang/srt/mem_cache/mamba_radix_cache.py

@@ -1052,7 +1052,7 @@ def _insert_helper(
             new_node = TreeNode()
             new_node.parent = node
             new_node.key = key
-            new_node.value = value
+            new_node.value = value.clone()
             new_node.mamba_value = mamba_value
             self.full_lru_list.insert_mru(new_node)
             self.mamba_lru_list.insert_mru(new_node)