diff --git a/tests/v1/structured_output/test_utils.py b/tests/v1/structured_output/test_utils.py index f5f8020220..3e4a12f7e5 100644 --- a/tests/v1/structured_output/test_utils.py +++ b/tests/v1/structured_output/test_utils.py @@ -5,9 +5,12 @@ import pytest from xgrammar import Grammar from xgrammar.testing import _is_grammar_accept_string +from vllm.sampling_params import SamplingParams, StructuredOutputsParams from vllm.v1.structured_output.backend_xgrammar import ( has_xgrammar_unsupported_json_features, + validate_xgrammar_grammar, ) +from vllm.v1.structured_output.utils import choice_as_grammar pytestmark = pytest.mark.cpu_test @@ -232,3 +235,46 @@ def test_pattern_properties_constrains_keys_and_values(pattern_properties_schema assert grammar_accepts(pattern_properties_schema, '{"grade_1": 5}') assert not grammar_accepts(pattern_properties_schema, '{"other": 5}') assert not grammar_accepts(pattern_properties_schema, '{"grade_1": "five"}') + + +class TestIsGrammarAcceptString: + class TestPR48115Regressions: + @pytest.mark.parametrize("codepoint", [*range(0x20), 0x7F]) + def test_choice_as_grammar_preserves_control_characters(self, codepoint): + choice = f"a{chr(codepoint)}f" + grammar = Grammar.from_ebnf(choice_as_grammar([choice])) + + assert _is_grammar_accept_string(grammar, choice) + assert not _is_grammar_accept_string(grammar, "a") + assert not _is_grammar_accept_string(grammar, "af") + assert not _is_grammar_accept_string(grammar, f"a\\u{codepoint:04x}f") + + @pytest.mark.parametrize( + ("choice", "other"), + [ + (r"a\nf", "a\nf"), + ('a quote " and a backslash \\', 'a quote " and a backslash '), + ("café", "cafe"), + ("日本語", "日本"), + ("😀", "😁"), + ], + ) + def test_choice_as_grammar_preserves_literal_choices(self, choice, other): + grammar = Grammar.from_ebnf(choice_as_grammar(["yes", choice])) + + assert _is_grammar_accept_string(grammar, "yes") + assert _is_grammar_accept_string(grammar, choice) + assert not _is_grammar_accept_string(grammar, other) + assert not _is_grammar_accept_string(grammar, choice + "extra") + + def test_validate_xgrammar_preserves_multiline_choice(self): + structured_outputs = StructuredOutputsParams(choice=["yes", "no\nplease"]) + validate_xgrammar_grammar( + SamplingParams(structured_outputs=structured_outputs) + ) + + assert structured_outputs.choice is None + grammar = Grammar.from_ebnf(structured_outputs.grammar) + assert _is_grammar_accept_string(grammar, "yes") + assert _is_grammar_accept_string(grammar, "no\nplease") + assert not _is_grammar_accept_string(grammar, r"no\nplease") diff --git a/vllm/v1/structured_output/utils.py b/vllm/v1/structured_output/utils.py index 4969789434..4bb438d6f3 100644 --- a/vllm/v1/structured_output/utils.py +++ b/vllm/v1/structured_output/utils.py @@ -567,9 +567,18 @@ def convert_lark_to_ebnf(grammar_str: str) -> str: def choice_as_grammar(choice: list[str]) -> str: def escape_ebnf_string(s: str) -> str: - """Escape special characters in a EBNF string.""" - # Escape double quotes and backslashes - return re.sub(r'(["\\])', r"\\\1", s) + """Escape EBNF literals, including raw LF, CR, and NUL terminators.""" + escapes = {"\\": r"\\", '"': r"\"", "\n": r"\n", "\r": r"\r", "\t": r"\t"} + + def escape_char(ch: str) -> str: + if ch in escapes: + return escapes[ch] + # Escape remaining C0 controls (U+0000-U+001F) and DEL (U+007F). + if ord(ch) < 0x20 or ord(ch) == 0x7F: + return f"\\u{ord(ch):04x}" + return ch + + return "".join(escape_char(ch) for ch in s) escaped_choices = (escape_ebnf_string(c) for c in choice) grammar = "root ::= " + " | ".join(f'"{c}"' for c in escaped_choices)