Method _dummy_run

fastdeploy/worker/xpu_model_runner.py:1405–1442

Use dummy inputs to run before formal execution. Args: num_tokens: Number of the input tokens batch_size: Batch size expected_decode_len: Expected decode length in_capturing: Is cuda graph in capturing state

(
        self,
        num_tokens: paddle.Tensor,
        batch_size: paddle.Tensor,
        expected_decode_len: int = 1,
        in_capturing: bool = False,
    )

            )

    def _dummy_run(
        self,
        num_tokens: paddle.Tensor,
        batch_size: paddle.Tensor,
        expected_decode_len: int = 1,
        in_capturing: bool = False,
    ) -> paddle.Tensor:
        """
        Use dummy inputs to run before formal execution.
        Args:
            num_tokens: Number of the input tokens
            batch_size: Batch size
            expected_decode_len: Expected decode length
            in_capturing: Is cuda graph in capturing state
        """
        input_length_list, max_dec_len_list, block_num = self.get_input_length_list(
            num_tokens=num_tokens,
            batch_size=batch_size,
            expected_decode_len=expected_decode_len,
        )
        self._dummy_prefill_inputs(
            input_length_list=input_length_list,
            max_dec_len_list=max_dec_len_list,
            block_num=block_num,
        )

        if self.speculative_method in ["mtp"]:
            self.proposer.dummy_prefill_inputs(
                num_tokens=num_tokens,
                batch_size=batch_size,
                expected_decode_len=1,
            )

        while True:
            self.execute_model(is_dummy_run=True, in_capturing=in_capturing)

            if int((self.share_inputs["seq_lens_this_time"] > 0).sum()) == 0:
                break

    def _init_speculative_proposer(self):
        """