Feat: Implement JSON Mode (response_format.type="json_object") (#4733)

Co-authored-by: Kyle Pena <kylepena@kyles-macbook-pro.turkey-marlin.ts.net>

python/sglang/srt/constrained/xgrammar_backend.py

@@ -158,6 +158,7 @@ class XGrammarGrammarBackend(BaseGrammarBackend):
     def dispatch_json(self, key_string: str) -> Optional[XGrammarGrammar]:
         try:
             if key_string == "$$ANY$$":
+                # Note: This builtin JSON grammar includes *all* valid JSON (including, for example, arrays at the root)
                 ctx = self.grammar_compiler.compile_builtin_json_grammar()
             else:
                 ctx = self.grammar_compiler.compile_json_schema(schema=key_string)

python/sglang/srt/openai_api/adapter.py

@@ -1105,6 +1105,8 @@ def v1_chat_generate_request(
             sampling_params["json_schema"] = convert_json_schema_to_str(
                 request.response_format.json_schema.schema_
             )
+        elif request.response_format and request.response_format.type == "json_object":
+            sampling_params["json_schema"] = '{"type": "object"}'
         elif (
             request.response_format and request.response_format.type == "structural_tag"
         ):