[V1] Fix json_object support with xgrammar (#15488)

Signed-off-by: Russell Bryant <rbryant@redhat.com>

requirements/common.txt

@@ -21,7 +21,7 @@ lm-format-enforcer >= 0.10.11, < 0.11
 llguidance >= 0.7.9, < 0.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64"
 outlines == 0.1.11
 lark == 1.2.2
-xgrammar == 0.1.16; platform_machine == "x86_64" or platform_machine == "aarch64"
+xgrammar == 0.1.17; platform_machine == "x86_64" or platform_machine == "aarch64"
 typing_extensions >= 4.10
 filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317
 partial-json-parser # used for parsing partial JSON outputs

tests/v1/entrypoints/llm/test_struct_output_generate.py

@@ -125,17 +125,9 @@ def test_structured_output(
             print(generated_text)
             assert generated_text is not None
 
-            # Parse to verify it is valid JSON
+            # Parse to verify it is a valid JSON object
             parsed_json = json.loads(generated_text)
-            allowed_types: tuple[type, ...] = (dict, )
-            if guided_decoding_backend.startswith("xgrammar"):
-                # TODO - we are currently too permissive with xgrammar and
-                # allow # any valid json (typically comes back as a list or
-                # object).  We can fix this by specifying a jsonschema of
-                # {"type": "object"}, # but we need this fix in a release
-                # first: https://github.com/mlc-ai/xgrammar/pull/264
-                allowed_types = (dict, list)
-            assert isinstance(parsed_json, allowed_types)
+            assert isinstance(parsed_json, dict)
 
     #
     # Test 3: test a jsonschema incompatible with xgrammar

vllm/model_executor/guided_decoding/__init__.py

@@ -79,12 +79,6 @@ def maybe_backend_fallback(
                     "xgrammar does not support Lark grammars and the "
                     "grammar failed to convert to GBNF.", "outlines")
 
-        elif guided_params.json_object:
-            # https://github.com/mlc-ai/xgrammar/issues/256
-            fallback_or_error(guided_params,
-                              "xgrammar does not support json_object.",
-                              "guidance")
-
         # If the xgrammar module cannot be imported successfully,
         # we should still allow users to use guided decoding with a fallback.
         elif not xgr_installed:

vllm/model_executor/guided_decoding/xgrammar_decoding.py

@@ -320,7 +320,10 @@ class XGrammarLogitsProcessor:
             elif self.config.grammar_str is not None:
                 self.ctx = compiler.compile_grammar(self.config.grammar_str)
             elif self.config.json_object:
-                self.ctx = compiler.compile_builtin_json_grammar()
+                any_whitespace = self.config.any_whitespace
+                self.ctx = compiler\
+                    .compile_json_schema('{"type": "object"}',
+                                         any_whitespace=any_whitespace)
             else:
                 raise ValueError(
                     "Invalid configuration for xgrammar logits processor")

vllm/v1/structured_output/backend_xgrammar.py

@@ -84,7 +84,9 @@ class XgrammarBackend(StructuredOutputBackend):
             ctx = self.compiler.compile_json_schema(
                 grammar_spec, any_whitespace=not self.disable_any_whitespace)
         elif request_type == StructuredOutputOptions.JSON_OBJECT:
-            ctx = self.compiler.compile_builtin_json_grammar()
+            ctx = self.compiler.compile_json_schema(
+                '{"type": "object"}',
+                any_whitespace=not self.disable_any_whitespace)
         elif request_type == StructuredOutputOptions.GRAMMAR:
             ctx = self.compiler.compile_grammar(grammar_spec)
         elif request_type == StructuredOutputOptions.REGEX: