[Qwen3-next] support mamba radix cache for overlap scheduler (#14792)

python/sglang/srt/managers/utils.py

@@ -24,7 +24,8 @@ class GenerationBatchResult:
     logits_output: Optional[LogitsProcessorOutput] = None
     pp_hidden_states_proxy_tensors: Optional[PPProxyTensors] = None
     next_token_ids: Optional[torch.Tensor] = None
-    num_accepted_tokens: Optional[int] = None
+    num_accepted_tokens: int = 0
+    accept_length_per_req_cpu: Optional[List[int]] = None
     can_run_cuda_graph: bool = False
 
     # For output processing