Skip to content

vllm.benchmarks.datasets.create_txt_slices_dataset

Convert a plain-text file (local path or URL) into a JSONL dataset compatible with CustomDataset (--dataset-name custom), by randomly slicing the tokenized text into prompts.

Each line of the output JSONL contains a prompt (decoded from a random slice of the tokenized source text) and an output_tokens count.

Usage

::

python -m vllm.benchmarks.datasets.create_txt_slices_dataset \
    --input  sonnet.txt \
    --output sonnet_dataset.jsonl \
    --tokenizer gpt2 \
    --num-prompts 1000 \
    --input-len 1024 \
    --output-len 128

The resulting JSONL file can then be used with the serving benchmark::

python -m vllm.benchmarks.serve \
    --dataset-name custom \
    --dataset-path sonnet_dataset.jsonl \
    ...

create_txt_slices_jsonl 

create_txt_slices_jsonl(
    *,
    input_path: str,
    output_path: str,
    tokenizer_name: str,
    num_prompts: int,
    input_len: int,
    output_len: int,
    range_ratio: RangeRatio = 0.0,
    seed: int = 0,
    trust_remote_code: bool = False,
) -> None

Read input_path, slice it into prompts, and write JSONL to output_path.

Source code in vllm/benchmarks/datasets/create_txt_slices_dataset.py 
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
def create_txt_slices_jsonl(
    *,
    input_path: str,
    output_path: str,
    tokenizer_name: str,
    num_prompts: int,
    input_len: int,
    output_len: int,
    range_ratio: RangeRatio = 0.0,
    seed: int = 0,
    trust_remote_code: bool = False,
) -> None:
    """Read *input_path*, slice it into prompts, and write JSONL to
    *output_path*."""

    tokenizer = AutoTokenizer.from_pretrained(
        tokenizer_name, trust_remote_code=trust_remote_code
    )

    text = load_text(input_path)
    if not text:
        raise ValueError("The text file is empty and cannot be sampled from.")

    token_ids = tokenizer(text, add_special_tokens=False).input_ids
    if not token_ids:
        raise ValueError("Tokenizing the text produced zero tokens; cannot sample.")

    rng_np = np.random.default_rng(seed)
    rng_py = random.Random(seed)

    input_lens, output_lens, _ = get_sampling_params(
        rng_np,
        num_prompts,
        range_ratio,
        input_len,
        output_len,
        tokenizer,
    )

    num_available_tokens = len(token_ids)

    records: list[dict[str, object]] = []
    for i in range(num_prompts):
        req_input_len = int(input_lens[i])
        req_output_len = int(output_lens[i])

        # Randomly select a start position and slice with cycling
        start_pos = rng_py.randint(0, num_available_tokens - 1)
        prompt_token_ids = [
            token_ids[(start_pos + j) % num_available_tokens]
            for j in range(req_input_len)
        ]
        prompt = tokenizer.decode(prompt_token_ids, skip_special_tokens=False)

        records.append({"prompt": prompt, "output_tokens": req_output_len})

    with open(output_path, "w", encoding="utf-8") as f:
        for record in records:
            f.write(json.dumps(record, ensure_ascii=False) + "\n")

    logger.info(
        "Wrote %d prompts to %s",
        len(records),
        output_path,
    )

load_text 

load_text(path: str) -> str

Load text from a local file or URL.

Source code in vllm/benchmarks/datasets/create_txt_slices_dataset.py 
47
48
49
50
51
52
53
def load_text(path: str) -> str:
    """Load text from a local file or URL."""
    if path.startswith(("http://", "https://")):
        with urllib.request.urlopen(path) as response:
            return response.read().decode("utf-8")
    with open(path, encoding="utf-8") as f:
        return f.read()