use '%llguidance' as marker to enable llg lark syntax

Author: mmoskal

common/CMakeLists.txt

@@ -96,7 +96,7 @@ if (LLAMA_LLGUIDANCE)
     set(LLGUIDANCE_PATH ${LLGUIDANCE_SRC}/target/release)
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        GIT_TAG 0cbe6b3a6ee72e5726c76f13fe67e21a5895f679
+        GIT_TAG d7b382916c466e6f26869671b8480229dc930423
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE

common/json-schema-to-grammar.cpp

@@ -992,7 +992,7 @@ class SchemaConverter {
 
 std::string json_schema_to_grammar(const json & schema) {
 #ifdef LLAMA_USE_LLGUIDANCE
-    return "llg:json:" + schema.dump();
+    return "%llguidance {}\nstart: %json " + schema.dump();
 #else
     return build_grammar([&](const llama_grammar_builder & callbacks) {
         auto copy = schema;

common/sampling.cpp

@@ -152,15 +152,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
     lparams.no_perf = params.no_perf;
 
     struct llama_sampler * grmr;
-    if (params.grammar.compare(0, 4, "llg:") == 0) {
+    if (params.grammar.compare(0, 11, "%llguidance") == 0) {
 #ifdef LLAMA_USE_LLGUIDANCE
-        auto gp = params.grammar.find(':', 4);
-        if (gp == std::string::npos) {
-            GGML_ABORT("invalid serialized grammar");
-        }
-        auto grm_type = params.grammar.substr(4, gp - 4);
-        auto grm_data = params.grammar.c_str() + gp + 1;
-        grmr = llama_sampler_init_llg(model, grm_type.c_str(), grm_data);
+        grmr = llama_sampler_init_llg(model, "lark", params.grammar.c_str());
 #else
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif