| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import pytest |
| 4 | |
| 5 | from vllm.sampling_params import RepetitionDetectionParams, SamplingParams |
| 6 | from vllm.v1.core.sched.utils import check_sequence_repetition, check_stop |
| 7 | from vllm.v1.request import Request, RequestStatus |
| 8 | |
| 9 | pytestmark = pytest.mark.cpu_test |
| 10 | |
| 11 | # ============================================================================ |
| 12 | # UNIT TESTS - check_sequence_repetition function |
| 13 | # ============================================================================ |
| 14 | |
| 15 | |
| 16 | class TestCheckSequenceRepetition: |
| 17 | """Unit tests for the check_sequence_repetition function.""" |
| 18 | |
| 19 | def test_simple_repetition_detected(self): |
| 20 | """Test detection of simple repetitive patterns.""" |
| 21 | token_ids = [1, 2, 3, 1, 2, 3, 1, 2, 3] |
| 22 | params = RepetitionDetectionParams( |
| 23 | max_pattern_size=3, |
| 24 | min_pattern_size=2, |
| 25 | min_count=3, |
| 26 | ) |
| 27 | assert check_sequence_repetition(token_ids, params) |
| 28 | |
| 29 | def test_repetition_below_min_count(self): |
| 30 | """Test that pattern below min_count is not detected.""" |
| 31 | token_ids = [1, 2, 3, 1, 2, 3] |
| 32 | params = RepetitionDetectionParams( |
| 33 | max_pattern_size=3, |
| 34 | min_pattern_size=2, |
| 35 | min_count=3, |
| 36 | ) |
| 37 | assert not check_sequence_repetition(token_ids, params) |
| 38 | |
| 39 | def test_two_token_pattern(self): |
| 40 | """Test detection of 2-token patterns.""" |
| 41 | token_ids = [1, 2, 1, 2, 1, 2, 1, 2] |
| 42 | params = RepetitionDetectionParams( |
| 43 | max_pattern_size=5, |
| 44 | min_pattern_size=2, |
| 45 | min_count=4, |
| 46 | ) |
| 47 | assert check_sequence_repetition(token_ids, params) |
| 48 | |
| 49 | def test_no_repetition_varied_sequence(self): |
| 50 | """Test that non-repetitive sequences are not flagged.""" |
| 51 | token_ids = [1, 2, 3, 4, 5, 6, 7, 8, 9] |
| 52 | params = RepetitionDetectionParams( |
| 53 | max_pattern_size=5, |
| 54 | min_pattern_size=2, |
| 55 | min_count=2, |
| 56 | ) |
| 57 | assert not check_sequence_repetition(token_ids, params) |
| 58 | |
| 59 | def test_partial_repetition_not_detected(self): |
| 60 | """Test that incomplete repetitions are not detected.""" |
| 61 | token_ids = [1, 2, 3, 1, 2, 3, 1, 2, 4] |
| 62 | params = RepetitionDetectionParams( |
| 63 | max_pattern_size=3, |
| 64 | min_pattern_size=2, |
| 65 | min_count=3, |
| 66 | ) |
| 67 | assert not check_sequence_repetition(token_ids, params) |
| 68 | |
| 69 | def test_empty_token_list(self): |
| 70 | """Test with empty token list.""" |
| 71 | params = RepetitionDetectionParams( |
| 72 | max_pattern_size=3, |
| 73 | min_pattern_size=2, |
| 74 | min_count=2, |
| 75 | ) |
| 76 | assert not check_sequence_repetition([], params) |
| 77 | |
| 78 | def test_detection_disabled_max_size_zero(self): |
| 79 | """Test that zero max_pattern_size disables detection.""" |
| 80 | token_ids = [1, 2, 1, 2, 1, 2] |
| 81 | params = RepetitionDetectionParams() |
| 82 | assert not check_sequence_repetition(token_ids, params) |
| 83 | |
| 84 | def test_invalid_min_count(self): |
| 85 | """Test that min_count < 2 returns False.""" |
| 86 | token_ids = [1, 2, 1, 2] |
| 87 | params = RepetitionDetectionParams() |
| 88 | assert not check_sequence_repetition(token_ids, params) |
| 89 | |
| 90 | def test_repetition_at_end_of_sequence(self): |
| 91 | """Test detection when repetition occurs at the end.""" |
| 92 | token_ids = [1, 2, 3, 4, 5, 6, 5, 6, 5, 6] |
| 93 | params = RepetitionDetectionParams( |
| 94 | max_pattern_size=3, |
| 95 | min_pattern_size=2, |
| 96 | min_count=3, |
| 97 | ) |
| 98 | assert check_sequence_repetition(token_ids, params) |
| 99 | |
| 100 | def test_large_pattern_many_repetitions(self): |
| 101 | """Test large pattern repeated many times.""" |
| 102 | token_ids = [1, 2, 3, 4, 5, 6, 7, 8] * 5 |
| 103 | params = RepetitionDetectionParams( |
| 104 | max_pattern_size=10, |
| 105 | min_pattern_size=2, |
| 106 | min_count=3, |
| 107 | ) |
| 108 | assert check_sequence_repetition(token_ids, params) |
| 109 | |
| 110 | |
| 111 | # ============================================================================ |
| 112 | # INTEGRATION TESTS - check_stop with repetition detection |
| 113 | # ============================================================================ |
| 114 | |
| 115 | |
| 116 | class TestRepetitionDetectionIntegration: |
| 117 | """Integration tests for repetition detection in check_stop.""" |
| 118 | |
| 119 | def test_basic_repetition_stops_generation(self): |
| 120 | """Test that repetition is detected and stops generation.""" |
| 121 | params = SamplingParams( |
| 122 | max_tokens=100, |
| 123 | repetition_detection=RepetitionDetectionParams( |
| 124 | max_pattern_size=5, |
| 125 | min_pattern_size=2, |
| 126 | min_count=3, |
| 127 | ), |
| 128 | ) |
| 129 | request = Request( |
| 130 | request_id="test", |
| 131 | prompt_token_ids=[1, 2, 3], |
| 132 | sampling_params=params, |
| 133 | pooling_params=None, |
| 134 | ) |
| 135 | request.append_output_token_ids([10, 20, 10, 20, 10, 20]) |
| 136 | assert check_stop(request, max_model_len=1024) |
| 137 | assert request.status == RequestStatus.FINISHED_REPETITION |
| 138 | assert request.stop_reason == "repetition_detected" |
| 139 | |
| 140 | def test_detection_disabled_no_stop(self): |
| 141 | """Test that disabled detection doesn't stop generation.""" |
| 142 | params = SamplingParams( |
| 143 | max_tokens=100, |
| 144 | ) |
| 145 | request = Request( |
| 146 | request_id="test", |
| 147 | prompt_token_ids=[1, 2, 3], |
| 148 | sampling_params=params, |
| 149 | pooling_params=None, |
| 150 | ) |
| 151 | request.append_output_token_ids([10, 20, 10, 20, 10, 20]) |
| 152 | assert not check_stop(request, max_model_len=1024) |
| 153 | |
| 154 | def test_repetition_respects_min_tokens(self): |
| 155 | """Test that repetition detection respects min_tokens.""" |
| 156 | params = SamplingParams( |
| 157 | min_tokens=10, |
| 158 | max_tokens=100, |
| 159 | repetition_detection=RepetitionDetectionParams( |
| 160 | max_pattern_size=5, |
| 161 | min_pattern_size=2, |
| 162 | min_count=3, |
| 163 | ), |
| 164 | ) |
| 165 | request = Request( |
| 166 | request_id="test", |
| 167 | prompt_token_ids=[1, 2, 3], |
| 168 | sampling_params=params, |
| 169 | pooling_params=None, |
| 170 | ) |
| 171 | request.append_output_token_ids([10, 20, 10, 20, 10, 20]) |
| 172 | assert not check_stop(request, max_model_len=1024) |
| 173 | |
| 174 | def test_no_repetition_continues_generation(self): |
| 175 | """Test that non-repetitive tokens don't stop generation.""" |
| 176 | params = SamplingParams( |
| 177 | max_tokens=100, |
| 178 | repetition_detection=RepetitionDetectionParams( |
| 179 | max_pattern_size=5, |
| 180 | min_pattern_size=2, |
| 181 | min_count=3, |
| 182 | ), |
| 183 | ) |
| 184 | request = Request( |
| 185 | request_id="test", |
| 186 | prompt_token_ids=[1, 2, 3], |
| 187 | sampling_params=params, |
| 188 | pooling_params=None, |
| 189 | ) |
| 190 | request.append_output_token_ids([10, 20, 30, 40, 50, 60]) |
| 191 | assert not check_stop(request, max_model_len=1024) |
| 192 | |
| 193 | def test_pattern_at_size_boundary(self): |
| 194 | """Test detection at exact pattern size boundary.""" |
| 195 | params = SamplingParams( |
| 196 | max_tokens=100, |
| 197 | repetition_detection=RepetitionDetectionParams( |
| 198 | max_pattern_size=3, |
| 199 | min_pattern_size=3, |
| 200 | min_count=2, |
| 201 | ), |
| 202 | ) |
| 203 | request = Request( |
| 204 | request_id="test", |
| 205 | prompt_token_ids=[1, 2], |
| 206 | sampling_params=params, |
| 207 | pooling_params=None, |
| 208 | ) |
| 209 | request.append_output_token_ids([10, 20, 30, 10, 20, 30]) |
| 210 | assert check_stop(request, max_model_len=1024) |
| 211 | assert request.status == RequestStatus.FINISHED_REPETITION |
| 212 | |
| 213 | def test_multiple_pattern_sizes_checked(self): |
| 214 | """Test that function checks pattern sizes in range.""" |
| 215 | params = SamplingParams( |
| 216 | max_tokens=100, |
| 217 | repetition_detection=RepetitionDetectionParams( |
| 218 | max_pattern_size=5, |
| 219 | min_pattern_size=2, |
| 220 | min_count=3, |
| 221 | ), |
| 222 | ) |
| 223 | request = Request( |
| 224 | request_id="test", |
| 225 | prompt_token_ids=[1], |
| 226 | sampling_params=params, |
| 227 | pooling_params=None, |
| 228 | ) |
| 229 | request.append_output_token_ids([7, 8, 9, 10, 7, 8, 9, 10, 7, 8, 9, 10]) |
| 230 | assert check_stop(request, max_model_len=1024) |
| 231 | assert request.status == RequestStatus.FINISHED_REPETITION |
| 232 | |
| 233 | def test_eos_takes_precedence_over_repetition(self): |
| 234 | """Test that EOS token stops before repetition check.""" |
| 235 | params = SamplingParams( |
| 236 | max_tokens=100, |
| 237 | stop_token_ids=[999], |
| 238 | repetition_detection=RepetitionDetectionParams( |
| 239 | max_pattern_size=5, |
| 240 | min_pattern_size=2, |
| 241 | min_count=3, |
| 242 | ), |
| 243 | ) |
| 244 | request = Request( |
| 245 | request_id="test", |
| 246 | prompt_token_ids=[1, 2, 3], |
| 247 | sampling_params=params, |
| 248 | pooling_params=None, |
| 249 | ) |
| 250 | request.append_output_token_ids([10, 20, 10, 20, 999]) |
| 251 | assert check_stop(request, max_model_len=1024) |
| 252 | assert request.status == RequestStatus.FINISHED_STOPPED |
| 253 | |
| 254 | def test_min_pattern_size_filters_small_patterns(self): |
| 255 | """Test that min_pattern_size filters out smaller patterns.""" |
| 256 | params = SamplingParams( |
| 257 | max_tokens=100, |
| 258 | repetition_detection=RepetitionDetectionParams( |
| 259 | max_pattern_size=5, |
| 260 | min_pattern_size=3, |
| 261 | min_count=3, |
| 262 | ), |
| 263 | ) |
| 264 | request = Request( |
| 265 | request_id="test", |
| 266 | prompt_token_ids=[1], |
| 267 | sampling_params=params, |
| 268 | pooling_params=None, |
| 269 | ) |
| 270 | request.append_output_token_ids([10, 20, 10, 20, 10, 20]) |
| 271 | assert not check_stop(request, max_model_len=1024) |
| 272 | |
| 273 | def test_high_repetition_threshold(self): |
| 274 | """Test that high min_count requires many repetitions.""" |
| 275 | params = SamplingParams( |
| 276 | max_tokens=100, |
| 277 | repetition_detection=RepetitionDetectionParams( |
| 278 | max_pattern_size=5, |
| 279 | min_pattern_size=2, |
| 280 | min_count=5, |
| 281 | ), |
| 282 | ) |
| 283 | request = Request( |
| 284 | request_id="test", |
| 285 | prompt_token_ids=[1], |
| 286 | sampling_params=params, |
| 287 | pooling_params=None, |
| 288 | ) |
| 289 | request.append_output_token_ids([10, 20, 10, 20, 10, 20]) |
| 290 | assert not check_stop(request, max_model_len=1024) |