| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import numpy as np |
| 5 | import pytest |
| 6 | |
| 7 | from vllm.benchmarks.datasets.utils import get_sampling_params |
| 8 | from vllm.tokenizers import TokenizerLike |
| 9 | |
| 10 | |
| 11 | class _FakeTokenizer(TokenizerLike): |
| 12 | """Minimal tokenizer implementing the TokenizerLike protocol |
| 13 | for testing get_sampling_params.""" |
| 14 | |
| 15 | def __init__(self, vocab_size: int = 1000, num_special_tokens: int = 0) -> None: |
| 16 | self._vocab_size = vocab_size |
| 17 | self._num_special_tokens = num_special_tokens |
| 18 | |
| 19 | # -- Properties required by TokenizerLike -- |
| 20 | |
| 21 | @classmethod |
| 22 | def from_pretrained(cls, path_or_repo_id, *a, **kw): |
| 23 | return cls() |
| 24 | |
| 25 | @property |
| 26 | def vocab_size(self) -> int: |
| 27 | return self._vocab_size |
| 28 | |
| 29 | @property |
| 30 | def all_special_tokens(self) -> list[str]: |
| 31 | return [] |
| 32 | |
| 33 | @property |
| 34 | def all_special_ids(self) -> list[int]: |
| 35 | return [] |
| 36 | |
| 37 | @property |
| 38 | def bos_token_id(self) -> int: |
| 39 | return 0 |
| 40 | |
| 41 | @property |
| 42 | def eos_token_id(self) -> int: |
| 43 | return 1 |
| 44 | |
| 45 | @property |
| 46 | def pad_token_id(self) -> int: |
| 47 | return 2 |
| 48 | |
| 49 | @property |
| 50 | def is_fast(self) -> bool: |
| 51 | return False |
| 52 | |
| 53 | @property |
| 54 | def max_token_id(self) -> int: |
| 55 | return self._vocab_size - 1 |
| 56 | |
| 57 | @property |
| 58 | def max_chars_per_token(self) -> int: |
| 59 | return 4 |
| 60 | |
| 61 | @property |
| 62 | def truncation_side(self) -> str: |
| 63 | return "right" |
| 64 | |
| 65 | def num_special_tokens_to_add(self) -> int: |
| 66 | return self._num_special_tokens |
| 67 | |
| 68 | def __call__(self, text, text_pair=None, **kw): # type: ignore[override] |
| 69 | raise NotImplementedError |
| 70 | |
| 71 | def get_vocab(self) -> dict[str, int]: |
| 72 | return {} |
| 73 | |
| 74 | def get_added_vocab(self) -> dict[str, int]: |
| 75 | return {} |
| 76 | |
| 77 | def encode(self, text, **kw) -> list[int]: # type: ignore[override] |
| 78 | raise NotImplementedError |
| 79 | |
| 80 | def apply_chat_template(self, messages, **kw): # type: ignore[override] |
| 81 | raise NotImplementedError |
| 82 | |
| 83 | def convert_tokens_to_ids(self, tokens): |
| 84 | raise NotImplementedError |
| 85 | |
| 86 | def convert_tokens_to_string(self, tokens: list[str]) -> str: |
| 87 | raise NotImplementedError |
| 88 | |
| 89 | def decode(self, ids, skip_special_tokens: bool = False) -> str: |
| 90 | raise NotImplementedError |
| 91 | |
| 92 | def convert_ids_to_tokens( |
| 93 | self, ids, skip_special_tokens: bool = False |
| 94 | ) -> list[str]: |
| 95 | raise NotImplementedError |
| 96 | |
| 97 | |
| 98 | class TestGetSamplingParams: |
| 99 | """Tests for ``get_sampling_params`` in ``vllm.benchmarks.datasets.shared``.""" |
| 100 | |
| 101 | # -- helpers -- |
| 102 | |
| 103 | @staticmethod |
| 104 | def _tok(vocab_size: int = 1000, num_special: int = 0) -> _FakeTokenizer: |
| 105 | return _FakeTokenizer(vocab_size=vocab_size, num_special_tokens=num_special) |
| 106 | |
| 107 | # -- return shape / dtype -- |
| 108 | |
| 109 | def test_returns_three_arrays(self): |
| 110 | rng = np.random.default_rng(0) |
| 111 | result = get_sampling_params(rng, 5, 0.0, 100, 50, self._tok()) |
| 112 | assert len(result) == 3 |
| 113 | for arr in result: |
| 114 | assert isinstance(arr, np.ndarray) |
| 115 | |
| 116 | @pytest.mark.parametrize("n", [1, 10, 100]) |
| 117 | def test_output_length_matches_num_requests(self, n: int): |
| 118 | rng = np.random.default_rng(42) |
| 119 | input_lens, output_lens, offsets = get_sampling_params( |
| 120 | rng, n, 0.0, 64, 32, self._tok() |
| 121 | ) |
| 122 | assert input_lens.shape == (n,) |
| 123 | assert output_lens.shape == (n,) |
| 124 | assert offsets.shape == (n,) |
| 125 | |
| 126 | # -- fixed lengths (range_ratio = 0) -- |
| 127 | |
| 128 | def test_zero_range_ratio_gives_constant_lengths(self): |
| 129 | rng = np.random.default_rng(7) |
| 130 | input_lens, output_lens, _ = get_sampling_params( |
| 131 | rng, 20, 0.0, 128, 64, self._tok() |
| 132 | ) |
| 133 | assert np.all(input_lens == 128) |
| 134 | assert np.all(output_lens == 64) |
| 135 | |
| 136 | def test_special_tokens_subtracted_from_input_only(self): |
| 137 | rng = np.random.default_rng(7) |
| 138 | input_lens, output_lens, _ = get_sampling_params( |
| 139 | rng, 10, 0.0, 100, 50, self._tok(num_special=4) |
| 140 | ) |
| 141 | # real_input_len = 100 - 4 = 96, range_ratio 0 → all 96 |
| 142 | assert np.all(input_lens == 96) |
| 143 | # special tokens are not subtracted from output length |
| 144 | assert np.all(output_lens == 50) |
| 145 | |
| 146 | # -- range ratios -- |
| 147 | |
| 148 | def test_input_range_bounds(self): |
| 149 | rng = np.random.default_rng(0) |
| 150 | ratio = 0.5 |
| 151 | base = 200 |
| 152 | input_lens, _, _ = get_sampling_params( |
| 153 | rng, 500, {"input": ratio, "output": 0.0}, base, 50, self._tok() |
| 154 | ) |
| 155 | lo = int(np.floor(base * (1 - ratio))) |
| 156 | hi = int(np.ceil(base * (1 + ratio))) |
| 157 | assert np.all(input_lens >= lo) |
| 158 | assert np.all(input_lens <= hi) |
| 159 | |
| 160 | def test_output_range_bounds(self): |
| 161 | rng = np.random.default_rng(0) |
| 162 | ratio = 0.3 |
| 163 | base = 100 |
| 164 | _, output_lens, _ = get_sampling_params( |
| 165 | rng, 500, {"input": 0.0, "output": ratio}, 50, base, self._tok() |
| 166 | ) |
| 167 | lo = max(1, int(np.floor(base * (1 - ratio)))) |
| 168 | hi = int(np.ceil(base * (1 + ratio))) |
| 169 | assert np.all(output_lens >= lo) |
| 170 | assert np.all(output_lens <= hi) |
| 171 | |
| 172 | def test_output_low_clamped_to_one(self): |
| 173 | """Even with a high ratio that would push output_low to 0, |
| 174 | the function clamps it to 1.""" |
| 175 | rng = np.random.default_rng(0) |
| 176 | # output_len=1, ratio=0.99 → floor(1*0.01)=0, should clamp to 1 |
| 177 | _, output_lens, _ = get_sampling_params( |
| 178 | rng, 50, {"input": 0.0, "output": 0.99}, 100, 1, self._tok() |
| 179 | ) |
| 180 | assert np.all(output_lens >= 1) |
| 181 | |
| 182 | # -- offsets bounded by vocab_size -- |
| 183 | |
| 184 | @pytest.mark.parametrize("vocab", [100, 32000, 128256]) |
| 185 | def test_offsets_within_vocab(self, vocab: int): |
| 186 | rng = np.random.default_rng(0) |
| 187 | _, _, offsets = get_sampling_params( |
| 188 | rng, 200, 0.0, 64, 32, self._tok(vocab_size=vocab) |
| 189 | ) |
| 190 | assert np.all(offsets >= 0) |
| 191 | assert np.all(offsets < vocab) |
| 192 | |
| 193 | # -- reproducibility -- |
| 194 | |
| 195 | def test_same_seed_same_results(self): |
| 196 | tok = self._tok() |
| 197 | rr = {"input": 0.3, "output": 0.2} |
| 198 | a = get_sampling_params(np.random.default_rng(42), 50, rr, 256, 64, tok) |
| 199 | b = get_sampling_params(np.random.default_rng(42), 50, rr, 256, 64, tok) |
| 200 | for arr_a, arr_b in zip(a, b): |
| 201 | np.testing.assert_array_equal(arr_a, arr_b) |
| 202 | |
| 203 | def test_different_seed_different_results(self): |
| 204 | tok = self._tok() |
| 205 | rr = {"input": 0.3, "output": 0.2} |
| 206 | a = get_sampling_params(np.random.default_rng(0), 50, rr, 256, 64, tok) |
| 207 | b = get_sampling_params(np.random.default_rng(1), 50, rr, 256, 64, tok) |
| 208 | # Extremely unlikely all three arrays match with different seeds |
| 209 | assert not all(np.array_equal(arr_a, arr_b) for arr_a, arr_b in zip(a, b)) |
| 210 | |
| 211 | # -- validation / error paths -- |
| 212 | |
| 213 | @pytest.mark.parametrize("bad_ratio", [-0.1, 1.0, 1.5]) |
| 214 | def test_invalid_input_range_ratio(self, bad_ratio: float): |
| 215 | rng = np.random.default_rng(0) |
| 216 | with pytest.raises(ValueError, match="input_range_ratio"): |
| 217 | get_sampling_params( |
| 218 | rng, 10, {"input": bad_ratio, "output": 0.0}, 100, 50, self._tok() |
| 219 | ) |
| 220 | |
| 221 | @pytest.mark.parametrize("bad_ratio", [-0.1, 1.0, 1.5]) |
| 222 | def test_invalid_output_range_ratio(self, bad_ratio: float): |
| 223 | rng = np.random.default_rng(0) |
| 224 | with pytest.raises(ValueError, match="output_range_ratio"): |
| 225 | get_sampling_params( |
| 226 | rng, 10, {"input": 0.0, "output": bad_ratio}, 100, 50, self._tok() |
| 227 | ) |
| 228 | |
| 229 | def test_invalid_dict_missing_keys(self): |
| 230 | rng = np.random.default_rng(0) |
| 231 | with pytest.raises(ValueError, match="input.*output"): |
| 232 | get_sampling_params(rng, 10, {"input": 0.1}, 100, 50, self._tok()) |
| 233 | |
| 234 | def test_input_len_zero_with_special_tokens(self): |
| 235 | """input_len < num_special_tokens → real_input_len = 0, which is fine |
| 236 | (range [0, 0]).""" |
| 237 | rng = np.random.default_rng(0) |
| 238 | input_lens, _, _ = get_sampling_params( |
| 239 | rng, 5, 0.0, 5, 50, self._tok(num_special=10) |
| 240 | ) |
| 241 | # real_input_len = max(0, 5 - 10) = 0 |
| 242 | assert np.all(input_lens == 0) |
| 243 | |
| 244 | # -- edge cases -- |
| 245 | |
| 246 | def test_single_request(self): |
| 247 | rng = np.random.default_rng(0) |
| 248 | i, o, off = get_sampling_params(rng, 1, 0.0, 100, 50, self._tok()) |
| 249 | assert i.shape == (1,) |
| 250 | assert o.shape == (1,) |
| 251 | assert off.shape == (1,) |
| 252 | |
| 253 | def test_large_num_requests(self): |
| 254 | rng = np.random.default_rng(0) |
| 255 | i, o, off = get_sampling_params(rng, 10_000, 0.5, 512, 128, self._tok()) |
| 256 | assert i.shape == (10_000,) |
| 257 | assert o.shape == (10_000,) |
| 258 | assert off.shape == (10_000,) |