Pytorch workflow Logits Processor problem

trtllm version: 0.20.0
GPU: 4090

class CombinedLogitsProcessor(LogitsProcessor):
    def __init__(self, stra) -> None:
        self.stra = stra
        self.step = 0

    def __call__(self, req_id: int, logits: torch.Tensor,
                token_ids: List[List[int]], stream_ptr: int,
                client_id: Optional[int]):
        print(self.step, len(token_ids[0]))
        self.step += 1

I wrote this processor, when disable_overlap_scheduler is False(default value),
It will always output the same length for step 0 and 1, and the current logit strategy will not cause effect to the next token, but to the next next token.