try to use llama.cpp's tool call parser first

Makefile

@@ -754,7 +754,7 @@ music_default.o: otherarch/acestep/music_adapter.cpp
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 # idiotic "for easier compilation"
-GPTTYPE_ADAPTER = gpttype_adapter.cpp otherarch/llama_v2.cpp otherarch/llama_v3.cpp src/llama.cpp src/llama-chat.cpp src/llama-mmap.cpp src/llama-context.cpp src/llama-adapter.cpp src/llama-arch.cpp src/llama-batch.cpp src/llama-vocab.cpp src/llama-grammar.cpp src/llama-sampler.cpp src/llama-kv-cache.cpp src/llama-kv-cache-iswa.cpp src/llama-memory-hybrid.cpp src/llama-memory-hybrid-iswa.cpp src/llama-memory-recurrent.cpp src/llama-model-loader.cpp src/llama-model.cpp src/llama-quant.cpp src/llama-hparams.cpp otherarch/gptj_v1.cpp otherarch/gptj_v2.cpp otherarch/gptj_v3.cpp otherarch/gpt2_v1.cpp otherarch/gpt2_v2.cpp otherarch/gpt2_v3.cpp otherarch/rwkv_v2.cpp otherarch/rwkv_v3.cpp otherarch/neox_v2.cpp otherarch/neox_v3.cpp otherarch/mpt_v3.cpp ggml/include/ggml.h ggml/include/ggml-cpu.h ggml/include/ggml-cuda.h include/llama.h otherarch/llama-util.h
+GPTTYPE_ADAPTER = gpttype_adapter.cpp common/chat.cpp otherarch/llama_v2.cpp otherarch/llama_v3.cpp src/llama.cpp src/llama-chat.cpp src/llama-mmap.cpp src/llama-context.cpp src/llama-adapter.cpp src/llama-arch.cpp src/llama-batch.cpp src/llama-vocab.cpp src/llama-grammar.cpp src/llama-sampler.cpp src/llama-kv-cache.cpp src/llama-kv-cache-iswa.cpp src/llama-memory-hybrid.cpp src/llama-memory-hybrid-iswa.cpp src/llama-memory-recurrent.cpp src/llama-model-loader.cpp src/llama-model.cpp src/llama-quant.cpp src/llama-hparams.cpp otherarch/gptj_v1.cpp otherarch/gptj_v2.cpp otherarch/gptj_v3.cpp otherarch/gpt2_v1.cpp otherarch/gpt2_v2.cpp otherarch/gpt2_v3.cpp otherarch/rwkv_v2.cpp otherarch/rwkv_v3.cpp otherarch/neox_v2.cpp otherarch/neox_v3.cpp otherarch/mpt_v3.cpp ggml/include/ggml.h ggml/include/ggml-cpu.h ggml/include/ggml-cuda.h include/llama.h otherarch/llama-util.h
 gpttype_adapter_failsafe.o: $(GPTTYPE_ADAPTER)
 	$(CXX) $(CXXFLAGS) $(FAILSAFE_FLAGS) -c $< -o $@
 gpttype_adapter.o: $(GPTTYPE_ADAPTER)

expose.cpp

@@ -347,6 +347,25 @@ extern "C"
         return chat_template.c_str();
     }
 
+    static std::string parsed_tool_calls = "";
+    const char* parse_chat_tool_calls(const char * generated_text,
+                                      const char * tools_json,
+                                      const char * chat_template,
+                                      const char * chat_template_kwargs_json,
+                                      const char * tool_choice,
+                                      bool parallel_tool_calls,
+                                      bool is_partial) {
+        parsed_tool_calls = gpttype_parse_chat_tool_calls(
+            generated_text ? generated_text : "",
+            tools_json ? tools_json : "",
+            chat_template ? chat_template : "",
+            chat_template_kwargs_json ? chat_template_kwargs_json : "",
+            tool_choice ? tool_choice : "",
+            parallel_tool_calls,
+            is_partial);
+        return parsed_tool_calls.c_str();
+    }
+
     const char* get_pending_output() {
        return gpttype_get_pending_output().c_str();
     }

gpttype_adapter.cpp

@@ -36,6 +36,7 @@
 #include "llama_v2.cpp"
 #include "llama_v3.cpp"
 #include "src/llama.cpp"
+#include "common/chat.cpp"
 #include "gptj_v1.cpp"
 #include "gptj_v2.cpp"
 #include "gptj_v3.cpp"
@@ -4725,6 +4726,95 @@ std::string gpttype_get_chat_template()
     return std::string(model_template.data(), model_template.size() - 1);
 }
 
+std::string gpttype_parse_chat_tool_calls(const std::string & generated_text,
+                                          const std::string & tools_json,
+                                          const std::string & chat_template,
+                                          const std::string & chat_template_kwargs_json,
+                                          const std::string & tool_choice,
+                                          bool parallel_tool_calls,
+                                          bool is_partial)
+{
+    try
+    {
+        if(generated_text.empty() || tools_json.empty())
+        {
+            return "";
+        }
+
+        json tools = json::parse(tools_json);
+        if(!tools.is_array() || tools.empty())
+        {
+            return "";
+        }
+
+        std::string tmpl_src = chat_template.empty() ? gpttype_get_chat_template() : chat_template;
+        if(tmpl_src.empty())
+        {
+            return "";
+        }
+
+        auto tmpls = common_chat_templates_init(nullptr, tmpl_src);
+
+        common_chat_msg msg;
+        msg.role = "user";
+        msg.content = "tool parser warmup";
+
+        common_chat_templates_inputs inputs;
+        inputs.use_jinja = true;
+        inputs.messages = { msg };
+        inputs.tools = common_chat_tools_parse_oaicompat(tools);
+        inputs.tool_choice = common_chat_tool_choice_parse_oaicompat(tool_choice.empty() ? "auto" : tool_choice);
+        inputs.parallel_tool_calls = parallel_tool_calls;
+        inputs.add_generation_prompt = true;
+
+        if(!chat_template_kwargs_json.empty())
+        {
+            json kwargs = json::parse(chat_template_kwargs_json);
+            if(kwargs.is_object())
+            {
+                for(const auto & item : kwargs.items())
+                {
+                    inputs.chat_template_kwargs[item.key()] = item.value().dump();
+                }
+            }
+        }
+
+        common_chat_params chat_params = common_chat_templates_apply(tmpls.get(), inputs);
+        if(chat_params.parser.empty())
+        {
+            return "";
+        }
+
+        common_chat_parser_params parser_params(chat_params);
+        parser_params.parse_tool_calls = true;
+        parser_params.parser.load(chat_params.parser);
+
+        common_chat_msg parsed = common_chat_parse(generated_text, is_partial, parser_params);
+        if(parsed.tool_calls.empty())
+        {
+            return "";
+        }
+
+        json tool_calls = parsed.to_json_oaicompat().value("tool_calls", json::array());
+        return tool_calls.dump();
+    }
+    catch(const std::exception & e)
+    {
+        if(debugmode == 1 && !is_quiet)
+        {
+            printf("\nNative tool parser failed: %s\n", e.what());
+        }
+    }
+    catch(...)
+    {
+        if(debugmode == 1 && !is_quiet)
+        {
+            printf("\nNative tool parser failed with unknown error.\n");
+        }
+    }
+    return "";
+}
+
 std::vector<int> gpttype_get_token_arr(const std::string & input, bool addbos)
 {
     std::vector<int> toks;

koboldcpp.py

@@ -968,6 +968,9 @@ def init_library():
     handle.token_count.restype = token_count_outputs
     handle.get_pending_output.restype = ctypes.c_char_p
     handle.get_chat_template.restype = ctypes.c_char_p
+    handle.parse_chat_tool_calls.argtypes = [ctypes.c_char_p, ctypes.c_char_p, ctypes.c_char_p, ctypes.c_char_p, ctypes.c_char_p, ctypes.c_bool, ctypes.c_bool]
+    handle.parse_chat_tool_calls.restype = ctypes.c_char_p
+
     handle.calc_new_state_kv.restype = ctypes.c_size_t
     handle.calc_new_state_tokencount.restype = ctypes.c_size_t
     handle.calc_old_state_kv.argtypes = [ctypes.c_int]
@@ -3667,6 +3670,47 @@ def toolcall_to_normalized_json(text,start_tag,end_tag,required_match_txt): #con
 
     return text #fallback
 
+def native_parse_toolcall_tags(text: str, genparams: dict) -> list:
+    global cached_chat_template, cached_jinja_kwargs
+    if not text or not genparams.get('using_openai_tools', False):
+        return []
+    tools = genparams.get('tools', [])
+    if not tools:
+        return []
+    try:
+        kwargs = dict(cached_jinja_kwargs or {})
+        user_kwargs = genparams.get("chat_template_kwargs")
+        if isinstance(user_kwargs, dict):
+            kwargs.update(user_kwargs)
+        if "reasoning_effort" in genparams and genparams["reasoning_effort"] is not None:
+            kwargs["reasoning_effort"] = genparams["reasoning_effort"]
+
+        tool_choice = genparams.get("tool_choice", "auto")
+        if isinstance(tool_choice, dict):
+            tool_choice = "required"
+        elif tool_choice is None:
+            tool_choice = "auto"
+        else:
+            tool_choice = str(tool_choice)
+
+        raw = handle.parse_chat_tool_calls(
+            text.encode("UTF-8"),
+            json.dumps(tools).encode("UTF-8"),
+            (cached_chat_template or "").encode("UTF-8"),
+            json.dumps(kwargs).encode("UTF-8"),
+            tool_choice.encode("UTF-8"),
+            True,
+            False)
+        if not raw:
+            return []
+        parsed = json.loads(ctypes.string_at(raw).decode("UTF-8", "ignore"))
+        if not isinstance(parsed, list):
+            return []
+        parsed = [normalize_tool_call_resp(obj) for obj in parsed if isinstance(obj, dict)]
+        return coerce_tool_argtypes(parsed, tools)
+    except Exception:
+        return []
+
 def repack_toolcall_tags(text: str, original_tools:list):
     global thinkformats, tool_call_pairs
     tool_calls = []
@@ -5168,8 +5212,11 @@ class KcppServerRequestHandler(http.server.SimpleHTTPRequestHandler):
         if api_format == 4 or api_format == 2 or api_format == 8 or api_format == 9:
             using_openai_tools = genparams.get('using_openai_tools', False)
             if using_openai_tools:
-                # first, check and potentially segment multiple tags for multi-tool calls
-                tool_calls = repack_toolcall_tags(recvtxt,genparams.get('tools', []))
+                # first, let llama.cpp's chat parser handle known template-specific tool formats
+                tool_calls = native_parse_toolcall_tags(recvtxt, genparams)
+                # fallback: check and potentially segment multiple tags for multi-tool calls
+                if not tool_calls:
+                    tool_calls = repack_toolcall_tags(recvtxt,genparams.get('tools', []))
                 if tool_calls and len(tool_calls)>0:
                     flat = []
                     for obj in tool_calls:

model_adapter.h

@@ -84,6 +84,13 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
 generation_outputs gpttype_generate(const generation_inputs inputs);
 bool gpttype_generate_abort();
 std::string gpttype_get_chat_template();
+std::string gpttype_parse_chat_tool_calls(const std::string & generated_text,
+                                          const std::string & tools_json,
+                                          const std::string & chat_template,
+                                          const std::string & chat_template_kwargs_json,
+                                          const std::string & tool_choice,
+                                          bool parallel_tool_calls,
+                                          bool is_partial);
 bool gpttype_batch_generate_enabled();
 int gpttype_batch_generate_submit(const generation_inputs inputs);
 bool gpttype_batch_generate_has_finished(int request_id);