Method _make_logprob_dict

fastdeploy/entrypoints/llm.py:484–517

Make a Logprob dictionary for a position. Args: logprobs: list of log probabilities logprob_token_ids: list of top token ids decoded_tokens: list of decoded top tokens rank: rank of the sampled token num_logprobs: number of logprobs requested

(
        logprobs: list[float],
        logprob_token_ids: list[int],
        decoded_tokens: Iterable[str | None],
        rank: int,
        num_logprobs: int,
    )

    @staticmethod
    def _make_logprob_dict(
        logprobs: list[float],
        logprob_token_ids: list[int],
        decoded_tokens: Iterable[str | None],
        rank: int,
        num_logprobs: int,
    ) -> dict[int, Logprob]:
        """Make a Logprob dictionary for a position.
        Args:
          logprobs: list of log probabilities
          logprob_token_ids: list of top token ids
          decoded_tokens: list of decoded top tokens
          rank: rank of the sampled token
          num_logprobs: number of logprobs requested
            by the user (in addition to sampled logprob)
        Returns:
          dict[token id, Logprob]
        """
        if num_logprobs == -1:
            num_logprobs = len(logprobs)
        # We do not need a special case for the sampled token
        # being in the topk, since inserting duplicated data
        # into a dictionary twice is the same as doing it once.
        topk_ranks = range(1, num_logprobs + 1)
        ranks = itertools.chain((rank,), topk_ranks)

        return {
            token_id: Logprob(
                logprob=logprob,
                rank=rank,
                decoded_token=token,
            )
            for token_id, logprob, rank, token in zip(logprob_token_ids, logprobs, ranks, decoded_tokens)
        }

    def _run_engine(
        self,