[Serve] Add interface for structural generation config

This PR exposes the options of grammar cache and compact json output in debug config.

Author: cyx-6

cpp/serve/config.cc

@@ -86,6 +86,8 @@ Result<DebugConfig> DebugConfig::FromJSON(const picojson::object& config) {
   } else {
     return TResult::Error("Uknown grammar execution mode " + grammar_execution_mode);
   }
+  res.disable_grammar_cache = json::LookupOrDefault<bool>(config, "disable_grammar_cache", false);
+  res.compact_json_output = json::LookupOrDefault<bool>(config, "compact_json_output", false);
   return TResult::Ok(res);
 }
 
@@ -114,6 +116,8 @@ picojson::object DebugConfig::AsJSON() const {
       break;
     }
   }
+  config["disable_grammar_cache"] = picojson::value(disable_grammar_cache);
+  config["compact_json_output"] = picojson::value(compact_json_output);
   return config;
 }
 

cpp/serve/config.h

@@ -63,6 +63,10 @@ class DebugConfig {
   SpecialRequestKind special_request = SpecialRequestKind::kNone;
   /*! \brief The grammar execution mode. */
   GrammarExecutionMode grammar_execution_mode = GrammarExecutionMode::kJumpForward;
+  /*! \brief Wether to use the grammar cache. */
+  bool disable_grammar_cache = false;
+  /*! \brief Wether to generate json with compact format style. */
+  bool compact_json_output = false;
 
   /*!
    * \brief Create debug config from JSON.

cpp/serve/engine.cc

@@ -495,7 +495,10 @@ class EngineImpl : public Engine {
 
     int n = request->generation_cfg->n;
     int rng_seed = request->generation_cfg->seed;
-    auto compiled_grammar = GetGrammarFromResponseFormat(request->generation_cfg->response_format);
+    auto compiled_grammar =
+        GetGrammarFromResponseFormat(request->generation_cfg->response_format,
+                                     request->generation_cfg->debug_config.disable_grammar_cache,
+                                     request->generation_cfg->debug_config.compact_json_output);
 
     std::vector<RequestStateEntry> rsentries;
     // Create the request state entry for the input.
@@ -814,14 +817,27 @@ class EngineImpl : public Engine {
   /*! \brief Create a grammar init context according to the response format. If the response format
    * is not JSON, return std::nullopt. */
   std::optional<xgrammar::CompiledGrammar> GetGrammarFromResponseFormat(
-      const ResponseFormat& response_format) {
+      const ResponseFormat& response_format, bool disable_grammar_cache, bool compact_json_output) {
     if (response_format.type != "json_object") {
       return std::nullopt;
     } else if (!response_format.schema) {
       return cached_grammar_compiler_.GetCompiledGrammarForJSON();
     } else {
-      return cached_grammar_compiler_.GetCompiledGrammarForJSONSchema(
-          response_format.schema.value());
+      std::optional<int> indent = std::nullopt;
+      std::optional<std::pair<std::string, std::string>> separators = std::nullopt;
+      if (!compact_json_output) {
+        std::optional<int> indent = 2;
+        std::optional<std::pair<std::string, std::string>> separators = std::make_pair(": ", ",");
+      }
+      if (disable_grammar_cache) {
+        return xgrammar::CompiledGrammar(
+            xgrammar::BuiltinGrammar::JSONSchema(response_format.schema.value(), indent, separators,
+                                                 true),
+            token_table_);
+      } else {
+        return cached_grammar_compiler_.GetCompiledGrammarForJSONSchema(
+            response_format.schema.value(), indent, separators, true);
+      }
     }
   }
 

python/mlc_llm/protocol/debug_protocol.py

@@ -17,6 +17,8 @@ class DebugConfig(BaseModel):
     pinned_system_prompt: bool = False
     special_request: Optional[Literal["query_engine_metrics"]] = None
     grammar_execution_mode: Literal["constraint", "jump_forward"] = "jump_forward"
+    disable_grammar_cache: bool = False
+    compact_json_output: bool = False
 
     """Special request indicators