vllm-engine-step

vllm/vllm/v1/engine/core.py · ef9339e · L710–740 · 固定 commit 上游

这是固定版本源码的只读摘录；可在线阅读，也可通过上游链接查看原文件。

710    def step(self) -> tuple[dict[int, EngineCoreOutputs], bool]:
711        """Schedule, execute, and make output.
712
713        Returns tuple of outputs and a flag indicating whether the model
714        was executed.
715        """
716        # Check for any requests remaining in the scheduler - unfinished,
717        # or finished and not yet removed from the batch.
718        if not self.scheduler.has_requests():
719            return {}, False
720        scheduler_output = self.scheduler.schedule(self._should_throttle_prefills())
721        future = self.model_executor.execute_model(scheduler_output, non_block=True)
722        grammar_output = self.scheduler.get_grammar_bitmask(scheduler_output)
723        with (
724            self.capture_iteration_details(scheduler_output) as iteration_details,
725            self.log_error_detail(scheduler_output),
726        ):
727            model_output = self._wait_for_output(future)
728            if model_output is None:
729                model_output = self.model_executor.sample_tokens(grammar_output)
730
731        # Before processing the model output, process any aborts that happened
732        # during the model execution.
733        self._process_aborts_queue()
734        engine_core_outputs = self.scheduler.update_from_output(
735            scheduler_output, model_output
736        )
737        self._attach_iteration_details(engine_core_outputs, iteration_details)
738        self._attach_weight_version(engine_core_outputs)
739
740        return engine_core_outputs, scheduler_output.total_num_scheduled_tokens > 0