diff --git a/tests/v1/structured_output/test_backend_xgrammar.py b/tests/v1/structured_output/test_backend_xgrammar.py index 339f0de17f..bc29112077 100644 --- a/tests/v1/structured_output/test_backend_xgrammar.py +++ b/tests/v1/structured_output/test_backend_xgrammar.py @@ -2,12 +2,18 @@ # SPDX-FileCopyrightText: Copyright contributors to the vLLM project +from typing import ClassVar + import pytest +from transformers import AutoTokenizer, PreTrainedTokenizerBase from xgrammar import Grammar from xgrammar.testing import _is_grammar_accept_string +from vllm.config import StructuredOutputsConfig, VllmConfig from vllm.sampling_params import SamplingParams, StructuredOutputsParams +from vllm.v1.structured_output.backend_types import StructuredOutputOptions from vllm.v1.structured_output.backend_xgrammar import ( + XgrammarBackend, has_xgrammar_unsupported_json_features, validate_xgrammar_grammar, ) @@ -462,3 +468,90 @@ class TestIsGrammarAcceptString: 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") + + +# ================================================ +# Test XgrammarBackend functionality +# ================================================ + + +class TestXgrammarBackend: + class TestPR58067Regressions: + """Here we check behavior logic of our flag + disable_any_whitespace when used with xgrammar backend. + """ + + _TOKENIZER = "openai-community/gpt2" + _VOCAB_SIZE = 50257 + tokenizer: ClassVar[PreTrainedTokenizerBase] + + @classmethod + @pytest.fixture(scope="class", autouse=True) + def _shared_tokenizer(cls): + cls.tokenizer = AutoTokenizer.from_pretrained(cls._TOKENIZER) + + @classmethod + def _backend(cls, *, disable_any_whitespace: bool) -> XgrammarBackend: + vllm_config = VllmConfig( + structured_outputs_config=StructuredOutputsConfig( + backend="xgrammar", + disable_any_whitespace=disable_any_whitespace, + ) + ) + return XgrammarBackend( + vllm_config, tokenizer=cls.tokenizer, vocab_size=cls._VOCAB_SIZE + ) + + @classmethod + def _accepts_json(cls, backend: XgrammarBackend, json_text: str) -> bool: + grammar = backend.compile_grammar(StructuredOutputOptions.JSON_OBJECT, "") + return grammar.accept_tokens("req", cls.tokenizer.encode(json_text)) + + def test_disable_any_whitespace_is_false(self): + """Verify expected behavior of disable_any_whitespace=False""" + backend = self._backend(disable_any_whitespace=False) + + # Check that grammar accepts json with NO space after colon + assert self._accepts_json(backend, '{"no_space_after_me":"yes_ofcourse"}') + + # Check that grammar accepts json with space after colon + assert self._accepts_json(backend, '{"yes_space_after_me": true}') + + # Check that grammar accepts json with NO space after comma + assert self._accepts_json( + backend, + '{"never_space_after_comma":"accepted","good_boy":":3"}', + ) + + # Check that grammar accepts json with space after comma + assert self._accepts_json( + backend, + '{"where_space_after_comma":42, "wow_thats_cool":4242}', + ) + + def test_disable_any_whitespace_is_true(self): + """Verify expected behavior of disable_any_whitespace=True""" + backend = self._backend(disable_any_whitespace=True) + + # Check that grammar accepts json with NO space after colon + assert self._accepts_json(backend, '{"no_space_after_me":"yes_ofcourse"}') + + # Check that grammar rejects json with space after colon + assert not self._accepts_json(backend, '{"yes_space_after_me": true}') + + # FIXME(arpera): + # Comma + separators=(",", ":") is wrong on xgrammar <=0.2.8 + # Seehttps://github.com/mlc-ai/xgrammar/issues/945 + # Re-enable both tests once vLLM pins xgrammar with the fix (>=0.2.9). + # + # # Check that grammar accepts json with NO space after comma + # assert self._accepts_json( + # backend, + # '{"never_space_after_comma":"accepted","good_boy":":3"}', + # ) + # + # # Check that grammar rejects json with space after comma + # assert not self._accepts_json( + # backend, + # '{"where_space_after_comma":42, "wow_thats_cool":4242}', + # ) diff --git a/vllm/v1/structured_output/backend_xgrammar.py b/vllm/v1/structured_output/backend_xgrammar.py index 733ed7db09..b675a9a3f1 100644 --- a/vllm/v1/structured_output/backend_xgrammar.py +++ b/vllm/v1/structured_output/backend_xgrammar.py @@ -89,13 +89,36 @@ class XgrammarBackend(StructuredOutputBackend): grammar_spec: str, stop_token_ids: set[int] | None = None, ) -> StructuredOutputGrammar: + # Note(arpera): + # Our flag disable_any_whitespace does NOT map directly to + # xgrammar's flag any_whitespace + # To achieve desired behavior of disable_any_whitespace + # we have to set not only any_whitespace + # but also specify a list of separators after which + # xgrammar must not insert spaces. + # This is a requirement of xgrammar's API, so we must comply with it. + # + # FIXME(arpera): + # Currently xgrammar v0.2.8 DOES emit spaces after comma + # even if we specify it in separators list. + # The bug has been reported to xgrammar team: + # https://github.com/mlc-ai/xgrammar/issues/945 + # Please, track that issue, and once it is resolved remove this comment. + # Upd. this bug was fixed in xgrammar main branch on Oct 8, 2026 + # and will be available in next release. + # So, remove this comment once xgrammar updates to v0.2.9 + separators = (",", ":") if self.disable_any_whitespace else None if request_type == StructuredOutputOptions.JSON: ctx = self.compiler.compile_json_schema( - grammar_spec, any_whitespace=not self.disable_any_whitespace + grammar_spec, + any_whitespace=not self.disable_any_whitespace, + separators=separators, ) elif request_type == StructuredOutputOptions.JSON_OBJECT: ctx = self.compiler.compile_json_schema( - '{"type": "object"}', any_whitespace=not self.disable_any_whitespace + '{"type": "object"}', + any_whitespace=not self.disable_any_whitespace, + separators=separators, ) elif request_type == StructuredOutputOptions.GRAMMAR: if grammar_is_likely_lark(grammar_spec):