Method _get_output_config

tensorrt_llm/sampling_params.py:486–521

(self, is_pytorch_backend: bool = False)

        return tllme.SamplingConfig(**llmapi_to_rt_param_map)

    def _get_output_config(self, is_pytorch_backend: bool = False) -> tllme.OutputConfig:
        sampling_param_fields = set(dir(SamplingParams))
        fields = [
            f
            for f in dir(tllme.OutputConfig)
            if not f.startswith("__") and f in sampling_param_fields
        ]

        config_kwargs = {f: getattr(self, f) for f in fields}

        if is_pytorch_backend:
            config_kwargs["return_log_probs"] = bool(self.logprobs)
            if self.prompt_logprobs and not self.return_context_logits:
                logger.info(
                    "Since prompt_logprobs is requested but return_context_logits is False, "
                    "internally enabling context logits for prompt logprobs computation. "
                    "context logits will be dropped after computation as the user didn't explicitly request them."
                )
                # TODO: Find a more elegant way to do this.
                # NOTE: This is an internal hack, so we can entirely avoid introducing
                # `prompt_logprobs` into the executor bindings and further into
                # model engine / sampler.
                # This is because, prompt_logprobs is a derived quantity from
                # context logits, and the capability to post-compute it
                # already exists in the worker. (see _get_logprobs in worker.py)
                config_kwargs["return_context_logits"] = True
        else:
            config_kwargs["return_log_probs"] = self._return_log_probs

        if config_kwargs.get("additional_model_outputs") is not None:
            config_kwargs["additional_model_outputs"] = [
                tllme.AdditionalModelOutput(name=output_name, gather_context=False)
                for output_name in config_kwargs["additional_model_outputs"]
            ]

        return tllme.OutputConfig(**config_kwargs)

    def _get_guided_decoding_params(self) -> tllme.GuidedDecodingParams:
        if self.guided_decoding is None: