Function load_requests

benchmark/benchmark_chat_completion.py:273–302

Load JSONL chat requests. Rows with list-type ``prompt`` (e.g. dapo-math-17k) are treated as message lists, matching ``benchmark_generate.py``. When ``tokenizer`` is provided (``--input-ids``), rows are converted to ``input_ids`` client-side.

(
    dataset_dir: str | Path | None = None,
    dataset_files: Sequence[str | Path] | None = None,
    datasets: Sequence[str] | None = None,
    num_prompts: int | None = None,
    shuffle: bool = False,
    seed: int = 1,
    tokenizer=None,
)

def load_requests(
    dataset_dir: str | Path | None = None,
    dataset_files: Sequence[str | Path] | None = None,
    datasets: Sequence[str] | None = None,
    num_prompts: int | None = None,
    shuffle: bool = False,
    seed: int = 1,
    tokenizer=None,
) -> list[BenchmarkRequest]:
    """Load JSONL chat requests.

    Rows with list-type ``prompt`` (e.g. dapo-math-17k) are treated as message lists,
    matching ``benchmark_generate.py``. When ``tokenizer`` is provided (``--input-ids``),
    rows are converted to ``input_ids`` client-side.
    """
    raw_rows = _read_raw_rows(
        dataset_dir=dataset_dir,
        dataset_files=dataset_files,
        datasets=datasets,
        num_prompts=num_prompts,
        shuffle=shuffle,
    )
    if shuffle:
        random.Random(seed).shuffle(raw_rows)
    if num_prompts is not None:
        raw_rows = raw_rows[:num_prompts]
    if not raw_rows:
        raise ValueError('No benchmark requests were loaded.')

    return [_normalize_row(row, dataset, row_index, tokenizer) for row, dataset, row_index in raw_rows]


def parse_sse_line(line: bytes | str) -> SSEEvent: