| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | """Check the watermarking goldens still reproduce.""" |
| 5 | |
| 6 | import dataclasses |
| 7 | from pathlib import Path |
| 8 | |
| 9 | import pytest |
| 10 | |
| 11 | from tests.watermarking.golden_candidates import ( |
| 12 | DETECTOR_FACTORIES, |
| 13 | GOLDEN_FLOAT_RTOL, |
| 14 | WATERMARK_CONFIG_FIELDS, |
| 15 | WATERMARKING_CANDIDATES, |
| 16 | compare_golden, |
| 17 | configured_algorithm_prf_combinations, |
| 18 | configured_algorithms, |
| 19 | read_goldens, |
| 20 | routing_boundary, |
| 21 | validate_golden_guards, |
| 22 | ) |
| 23 | from vllm.config.watermarking import WatermarkConfig |
| 24 | |
| 25 | GOLDENS_PATH = Path(__file__).with_name("watermarking_goldens.json") |
| 26 | |
| 27 | CANDIDATES_BY_ID = {candidate.id: candidate for candidate in WATERMARKING_CANDIDATES} |
| 28 | |
| 29 | NEGATIVE_CANDIDATE_ID = "gumbel-philox-key42-cw4-wrong-key" |
| 30 | |
| 31 | |
| 32 | @pytest.fixture(scope="module") |
| 33 | def goldens() -> dict[str, dict]: |
| 34 | return read_goldens(GOLDENS_PATH) |
| 35 | |
| 36 | |
| 37 | def test_goldens_cover_every_candidate(goldens: dict[str, dict]): |
| 38 | candidate_ids = set(CANDIDATES_BY_ID) |
| 39 | |
| 40 | assert len(candidate_ids) == len(WATERMARKING_CANDIDATES) |
| 41 | assert set(goldens) == candidate_ids, "golden contract does not match candidates" |
| 42 | candidate_combinations = { |
| 43 | (candidate.scheme, candidate.prf) for candidate in WATERMARKING_CANDIDATES |
| 44 | } |
| 45 | assert candidate_combinations == configured_algorithm_prf_combinations() |
| 46 | assert set(DETECTOR_FACTORIES) == configured_algorithms() |
| 47 | |
| 48 | |
| 49 | def test_golden_fixture_guards(goldens: dict[str, dict]): |
| 50 | validate_golden_guards(goldens) |
| 51 | |
| 52 | |
| 53 | def test_resolved_records_every_watermark_config_field(): |
| 54 | assert {field.name for field in dataclasses.fields(WatermarkConfig)} == ( |
| 55 | set(WATERMARK_CONFIG_FIELDS) | {"key"} |
| 56 | ) |
| 57 | |
| 58 | |
| 59 | @pytest.mark.parametrize( |
| 60 | "candidate", |
| 61 | WATERMARKING_CANDIDATES, |
| 62 | ids=lambda candidate: candidate.id, |
| 63 | ) |
| 64 | def test_watermarking_candidate_golden(candidate, goldens: dict[str, dict]): |
| 65 | differences = compare_golden(candidate, goldens[candidate.id]) |
| 66 | |
| 67 | assert not differences, "\n".join( |
| 68 | [ |
| 69 | f"{candidate.id}: golden drift " |
| 70 | f"(tolerance {GOLDEN_FLOAT_RTOL:g} on score and p_value)", |
| 71 | *differences, |
| 72 | ] |
| 73 | ) |
| 74 | |
| 75 | |
| 76 | def test_negative_row_is_not_watermarked(goldens: dict[str, dict]): |
| 77 | candidate = CANDIDATES_BY_ID[NEGATIVE_CANDIDATE_ID] |
| 78 | |
| 79 | assert candidate.detection_key != candidate.key |
| 80 | assert goldens[NEGATIVE_CANDIDATE_ID]["detection"]["is_watermarked"] is False |
| 81 | |
| 82 | |
| 83 | @pytest.mark.parametrize( |
| 84 | "candidate", |
| 85 | [ |
| 86 | candidate |
| 87 | for candidate in WATERMARKING_CANDIDATES |
| 88 | if candidate.scheme == "dual_key_gumbel" |
| 89 | and 0 < candidate.scheme_config.generation_alpha < 1 |
| 90 | ], |
| 91 | ids=lambda candidate: candidate.id, |
| 92 | ) |
| 93 | def test_routing_sequence_resolves_alpha(candidate): |
| 94 | alpha = candidate.scheme_config.generation_alpha |
| 95 | uniforms = candidate.fixture.routing_uniforms |
| 96 | boundary = routing_boundary(alpha) |
| 97 | share = sum(1 for uniform in uniforms if boundary <= uniform) / len(uniforms) |
| 98 | |
| 99 | assert abs(share - alpha) <= 0.03 |