Remove toolcall dependency from common

Author: bandoti

common/CMakeLists.txt

@@ -139,7 +139,3 @@ endif ()
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
-
-if (LLAMA_TOOLCALL)
-    target_link_libraries(${TARGET} PUBLIC toolcall)
-endif()

common/arg.cpp

@@ -2143,24 +2143,22 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
 
-#ifdef LLAMA_USE_TOOLCALL
     add_opt(common_arg(
         {"--tools"}, "JINJA_TOOLS",
         "set to URI of a Model Context Protocol server, or "
         "a JSON array containing tool definitions (requires --jinja)",
         [](common_params &params, const std::string & value) {
-            params.jinja_tools.tools(value);
+            params.toolcall.tools = value;
 
         }).set_examples({LLAMA_EXAMPLE_MAIN}));
 
     add_opt(common_arg(
         {"--tool-choice"}, "JINJA_TOOL_CHOICE",
         "set to \"auto\", \"required\", or \"none\" (default: \"auto\")",
         [](common_params &params, const std::string & value) {
-            params.jinja_tools.choice(value);
+            params.toolcall.choice = value;
 
         }).set_examples({LLAMA_EXAMPLE_MAIN}));
-#endif
 
     add_opt(common_arg(
         {"-sps", "--slot-prompt-similarity"}, "SIMILARITY",

common/common.h

@@ -4,10 +4,6 @@
 
 #include "llama-cpp.h"
 
-#ifdef LLAMA_USE_TOOLCALL
-#    include "toolcall-params.h"
-#endif
-
 #include <set>
 #include <string>
 #include <vector>
@@ -212,6 +208,11 @@ enum common_reasoning_format {
     COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
 };
 
+struct common_toolcall_params {
+    std::string tools  = "";
+    std::string choice = "auto";
+};
+
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =  4096; // context size
@@ -358,9 +359,7 @@ struct common_params {
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
 
-#ifdef LLAMA_USE_TOOLCALL
-    toolcall::params jinja_tools;
-#endif
+    struct common_toolcall_params toolcall;
 
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
 

examples/main/CMakeLists.txt

@@ -2,4 +2,9 @@ set(TARGET llama-cli)
 add_executable(${TARGET} main.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+
+if (LLAMA_TOOLCALL)
+    target_link_libraries(${TARGET} PRIVATE toolcall)
+endif()
+
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/main/main.cpp

@@ -155,6 +155,11 @@ int main(int argc, char ** argv) {
 
     auto & sparams = params.sampling;
 
+#ifdef LLAMA_USE_TOOLCALL
+    // Ensure parameters are validated before the model loads
+    toolcall::params tc_params(params.toolcall.tools, params.toolcall.choice);
+#endif
+
     // save choice to use color for later
     // (note for later: this is a slightly awkward choice)
     console::init(params.simple_io, params.use_color);
@@ -323,7 +328,7 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd_inp;
 
 #ifdef LLAMA_USE_TOOLCALL
-    auto tc_handler = toolcall::create_handler(params.jinja_tools);
+    auto tc_handler = toolcall::create_handler(tc_params);
     if (tc_handler) {
         tc_handler->initialize();
     }