| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Guidance backend must not choke on speculative -1 padding. |
| 4 | |
| 5 | Speculative drafters (e.g. ngram_gpu) pad unfilled draft slots with a -1 |
| 6 | sentinel. llguidance raises ``OverflowError`` on a negative id, so |
| 7 | ``GuidanceGrammar.validate_tokens`` must stop at the first sentinel. The mock |
| 8 | matcher below emulates that raise, so each test fails on the unpatched backend |
| 9 | and passes with the guard. |
| 10 | """ |
| 11 | |
| 12 | import pytest |
| 13 | |
| 14 | from vllm.v1.structured_output.backend_guidance import GuidanceGrammar |
| 15 | |
| 16 | pytestmark = pytest.mark.skip_global_cleanup |
| 17 | |
| 18 | |
| 19 | class _Matcher: |
| 20 | """Stand-in for ``llguidance.LLMatcher`` that rejects negative ids.""" |
| 21 | |
| 22 | def __init__(self) -> None: |
| 23 | self.tokens: list[int] | None = None |
| 24 | |
| 25 | def is_stopped(self) -> bool: |
| 26 | return False |
| 27 | |
| 28 | def validate_tokens(self, tokens: list[int]) -> int: |
| 29 | if any(t < 0 for t in tokens): |
| 30 | raise OverflowError("llguidance rejects negative token ids") |
| 31 | self.tokens = tokens |
| 32 | return len(tokens) |
| 33 | |
| 34 | def get_error(self): |
| 35 | return None |
| 36 | |
| 37 | |
| 38 | class _Tokenizer: |
| 39 | eos_token = 0 |
| 40 | |
| 41 | |
| 42 | def _grammar(matcher: _Matcher) -> GuidanceGrammar: |
| 43 | return GuidanceGrammar(ll_matcher=matcher, ll_tokenizer=_Tokenizer(), vocab_size=16) |
| 44 | |
| 45 | |
| 46 | def test_validate_tokens_stops_at_trailing_padding() -> None: |
| 47 | matcher = _Matcher() |
| 48 | assert _grammar(matcher).validate_tokens([3, 4, -1, -1]) == [3, 4] |
| 49 | assert matcher.tokens == [3, 4] |
| 50 | |
| 51 | |
| 52 | def test_validate_tokens_single_negative_does_not_raise() -> None: |
| 53 | matcher = _Matcher() |
| 54 | assert _grammar(matcher).validate_tokens([-1]) == [] |
| 55 | assert matcher.tokens is None |
| 56 | |
| 57 | |
| 58 | def test_validate_tokens_all_padding_returns_empty() -> None: |
| 59 | matcher = _Matcher() |
| 60 | assert _grammar(matcher).validate_tokens([-1, -1]) == [] |
| 61 | assert matcher.tokens is None |