Reduce overhead when fork(1) (#375)

python/sglang/lang/interpreter.py

@@ -256,9 +256,15 @@ class StreamExecutor:
         ret = self.meta_info.get(name, None)
         return ret
 
-    def fork(self, number: int, position_ids_offset: Optional[List[int]] = None):
-        self.submit(SglCommitLazy())
-        self.sync()
+    def fork(
+        self,
+        number: int,
+        position_ids_offset: Optional[List[int]] = None,
+        copy: bool = False,
+    ):
+        if number > 1 or copy:
+            self.submit(SglCommitLazy())
+            self.sync()
 
         number = int(number)
 
@@ -641,15 +647,20 @@ class ProgramState:
         yield
         self.stream_executor.submit(SglVarScopeEnd(name))
 
-    def fork(self, number: int = 1, position_ids_offset: Optional[List[int]] = None):
-        stream_executors = self.stream_executor.fork(number, position_ids_offset)
+    def fork(
+        self,
+        number: int = 1,
+        position_ids_offset: Optional[List[int]] = None,
+        copy: bool = False,
+    ):
+        stream_executors = self.stream_executor.fork(number, position_ids_offset, copy)
         states = [ProgramState(x) for x in stream_executors]
         state_group = ProgramStateGroup(states, self)
         return state_group
 
     @contextmanager
     def copy(self, position_ids_offset: Optional[List[int]] = None):
-        state_group = self.fork(1, position_ids_offset)
+        state_group = self.fork(1, position_ids_offset, True)
         try:
             yield state_group[0]
         finally:

python/sglang/srt/managers/router/radix_cache.py

@@ -179,7 +179,9 @@ class RadixCache:
 
     def _print_helper(self, node, indent):
         for _, child in node.children.items():
-            print(" " * indent, len(child.key), child.key[:10], f"r={child.ref_counter}")
+            print(
+                " " * indent, len(child.key), child.key[:10], f"r={child.ref_counter}"
+            )
             self._print_helper(child, indent=indent + 2)
 
     def _delete_leaf(self, node):