| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Shared utilities for benchmark dataset sampling.""" |
| 4 | |
| 5 | import logging |
| 6 | import math |
| 7 | |
| 8 | import numpy as np |
| 9 | |
| 10 | from vllm.tokenizers import TokenizerLike |
| 11 | |
| 12 | logger = logging.getLogger(__name__) |
| 13 | |
| 14 | # Type alias: a single float applies to both ISL and OSL; a dict allows |
| 15 | # specifying them independently via ``{"input": …, "output": …}``. |
| 16 | RangeRatio = float | dict[str, float] |
| 17 | |
| 18 | |
| 19 | def _resolve_range_ratios( |
| 20 | range_ratio: RangeRatio, |
| 21 | ) -> tuple[float, float]: |
| 22 | """Return ``(input_range_ratio, output_range_ratio)`` from *range_ratio*. |
| 23 | |
| 24 | *range_ratio* is either a single float (used for both input and output) |
| 25 | or a dict with ``"input"`` and ``"output"`` keys. |
| 26 | """ |
| 27 | if isinstance(range_ratio, dict): |
| 28 | try: |
| 29 | return float(range_ratio["input"]), float(range_ratio["output"]) |
| 30 | except KeyError as exc: |
| 31 | raise ValueError( |
| 32 | "When range_ratio is a dict it must contain 'input' and " |
| 33 | f"'output' keys, got: {sorted(range_ratio)}" |
| 34 | ) from exc |
| 35 | ratio = float(range_ratio) |
| 36 | return ratio, ratio |
| 37 | |
| 38 | |
| 39 | def get_sampling_params( |
| 40 | rng: np.random.Generator, |
| 41 | num_requests: int, |
| 42 | range_ratio: RangeRatio, |
| 43 | input_len: int, |
| 44 | output_len: int, |
| 45 | tokenizer: TokenizerLike, |
| 46 | ) -> tuple[np.ndarray, np.ndarray, np.ndarray]: |
| 47 | """Sample per-request input/output token lengths and vocab offsets. |
| 48 | |
| 49 | Lengths are drawn uniformly from integer ranges around the configured |
| 50 | means, controlled by *range_ratio*. It may be a single ``float`` |
| 51 | (applied to both input and output) or a ``dict`` with ``"input"`` and |
| 52 | ``"output"`` keys for independent control. |
| 53 | |
| 54 | Tokenizer special tokens are subtracted from ``input_len`` before |
| 55 | computing the sampling interval. |
| 56 | |
| 57 | Returns: |
| 58 | (input_lens, output_lens, offsets) – three 1-D ``np.ndarray`` of |
| 59 | shape ``(num_requests,)``. |
| 60 | |
| 61 | """ |
| 62 | input_range_ratio, output_range_ratio = _resolve_range_ratios(range_ratio) |
| 63 | |
| 64 | if not (0.0 <= input_range_ratio < 1.0): |
| 65 | raise ValueError("input_range_ratio must be in [0, 1).") |
| 66 | if not (0.0 <= output_range_ratio < 1.0): |
| 67 | raise ValueError("output_range_ratio must be in [0, 1).") |
| 68 | num_special_tokens = int(tokenizer.num_special_tokens_to_add()) |
| 69 | real_input_len = max(0, int(input_len) - num_special_tokens) |
| 70 | input_low = math.floor(real_input_len * (1 - input_range_ratio)) |
| 71 | input_high = math.ceil(real_input_len * (1 + input_range_ratio)) |
| 72 | output_low = math.floor(output_len * (1 - output_range_ratio)) |
| 73 | output_high = math.ceil(output_len * (1 + output_range_ratio)) |
| 74 | # Ensure the lower bound for output length is at least 1 to |
| 75 | # prevent sampling 0 tokens. |
| 76 | output_low = max(output_low, 1) |
| 77 | output_high = max(output_high, 1) |
| 78 | |
| 79 | if input_low > input_high: |
| 80 | raise ValueError( |
| 81 | f"Invalid input sampling interval: low={input_low} > high={input_high}" |
| 82 | ) |
| 83 | if output_low > output_high: |
| 84 | raise ValueError( |
| 85 | f"Invalid output sampling interval: low={output_low} > high={output_high}" |
| 86 | ) |
| 87 | |
| 88 | logger.info( |
| 89 | "Sampling input_len from [%s, %s] and output_len from [%s, %s]", |
| 90 | input_low, |
| 91 | input_high, |
| 92 | output_low, |
| 93 | output_high, |
| 94 | ) |
| 95 | |
| 96 | input_lens = rng.integers(input_low, input_high + 1, size=num_requests) |
| 97 | output_lens = rng.integers(output_low, output_high + 1, size=num_requests) |
| 98 | offsets = rng.integers(0, tokenizer.vocab_size, size=num_requests) |
| 99 | return input_lens, output_lens, offsets |