Rename: common/chat.*, common_chat_{inputs -> params}

ochafik · ochafik · commit ed7c622d7897 · 2025-01-29T21:18:49.000Z
diff --git a/Makefile b/Makefile
@@ -52,7 +52,7 @@ TEST_TARGETS = \
 	tests/test-arg-parser \
 	tests/test-autorelease \
 	tests/test-backend-ops \
-	tests/test-chat-handler \
+	tests/test-chat \
 	tests/test-chat-template \
 	tests/test-double-float \
 	tests/test-grammar-integration \
@@ -984,7 +984,7 @@ OBJ_COMMON = \
 	$(DIR_COMMON)/ngram-cache.o \
 	$(DIR_COMMON)/sampling.o \
 	$(DIR_COMMON)/speculative.o \
-	$(DIR_COMMON)/chat-handler.o \
+	$(DIR_COMMON)/chat.o \
 	$(DIR_COMMON)/build-info.o \
 	$(DIR_COMMON)/json-schema-to-grammar.o
 
@@ -1363,8 +1363,8 @@ llama-server: \
 	examples/server/httplib.h \
 	examples/server/index.html.hpp \
 	examples/server/loading.html.hpp \
-	common/chat-handler.cpp \
-	common/chat-handler.hpp \
+	common/chat.cpp \
+	common/chat.hpp \
 	common/chat-template.hpp \
 	common/json.hpp \
 	common/minja.hpp \
@@ -1475,7 +1475,7 @@ tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \
 	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-chat-handler: tests/test-chat-handler.cpp \
+tests/test-chat: tests/test-chat.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt
@@ -56,8 +56,8 @@ add_library(${TARGET} STATIC
     arg.cpp
     arg.h
     base64.hpp
-    chat-handler.cpp
-    chat-handler.hpp
+    chat.cpp
+    chat.hpp
     chat-template.hpp
     common.cpp
     common.h
diff --git a/common/chat.cpp b/common/chat.cpp
diff --git a/common/chat.hpp b/common/chat.hpp
@@ -1,11 +1,5 @@
-/*
-    Copyright 2024 Google LLC
+// Chat support (incl. tool call grammar constraining & output parsing) w/ generic & custom template handlers.
 
-    Use of this source code is governed by an MIT-style
-    license that can be found in the LICENSE file or at
-    https://opensource.org/licenses/MIT.
-*/
-// SPDX-License-Identifier: MIT
 #pragma once
 
 #include "common.h"
@@ -16,7 +10,7 @@
 
 using json = nlohmann::ordered_json;
 
-struct common_chat_params {
+struct common_chat_inputs {
     json messages;
     json tools;
     json tool_choice;
@@ -29,7 +23,7 @@ struct common_chat_params {
 
 typedef std::function<common_chat_msg(const std::string & input)> common_chat_parser;
 
-struct common_chat_data {
+struct common_chat_params {
     json prompt;
     std::string grammar;
     std::vector<common_grammar_trigger> grammar_triggers;
@@ -39,4 +33,4 @@ struct common_chat_data {
     bool grammar_lazy = false;
 };
 
-struct common_chat_data common_chat_init(const common_chat_template & tmpl, const struct common_chat_params & params);
+struct common_chat_params common_chat_params_init(const common_chat_template & tmpl, const struct common_chat_inputs & params);
diff --git a/common/common.cpp b/common/common.cpp
@@ -12,7 +12,7 @@
 #include "json.hpp"
 #include "json-schema-to-grammar.h"
 #include "llama.h"
-#include "chat-handler.hpp"
+#include "chat.hpp"
 #include "chat-template.hpp"
 
 #include <algorithm>
@@ -1776,12 +1776,12 @@ bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {
     if (use_jinja) {
         try {
             auto chat_template = common_chat_template(tmpl, "<s>", "</s>");
-            common_chat_params params;
+            common_chat_inputs params;
             params.messages = json::array({{
                 {"role", "user"},
                 {"content", "test"},
             }});
-            common_chat_init(chat_template, params);
+            common_chat_params_init(chat_template, params);
             return true;
         } catch (const std::exception & e) {
             LOG_ERR("%s: failed to apply template: %s\n", __func__, e.what());
@@ -1803,10 +1803,10 @@ std::string common_chat_apply_template(
         for (const auto & msg : msgs) {
             messages.push_back({{"role", msg.role}, {"content", msg.content}});
         }
-        common_chat_params params;
+        common_chat_inputs params;
         params.messages = messages;
         params.add_generation_prompt = add_ass;
-        auto data = common_chat_init(tmpl, params);
+        auto data = common_chat_params_init(tmpl, params);
         return data.prompt;
     }
 
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
@@ -1824,16 +1824,16 @@ struct server_context {
 
         if (use_jinja) {
             auto templates = common_chat_templates_from_model(model, "");
-            common_chat_params params;
+            common_chat_inputs params;
             params.messages = json::array({{
                 {"role", "user"},
                 {"content", "test"},
             }});
             GGML_ASSERT(templates.template_default);
             try {
-                common_chat_init(*templates.template_default, params);
+                common_chat_params_init(*templates.template_default, params);
                 if (templates.template_tool_use) {
-                    common_chat_init(*templates.template_tool_use, params);
+                    common_chat_params_init(*templates.template_tool_use, params);
                 }
                 return true;
             } catch (const std::exception & e) {
@@ -3787,10 +3787,10 @@ int main(int argc, char ** argv) {
         std::vector<server_task> tasks;
 
         try {
-            common_chat_data chat_data;
+            common_chat_params chat_data;
             bool add_special = false;
             if (tmpl && ctx_server.params_base.use_jinja) {
-                chat_data = common_chat_init(*tmpl, {
+                chat_data = common_chat_params_init(*tmpl, {
                     /* .messages = */ json_value(data, "messages", json::array()),
                     /* .tools = */   json_value(data, "tools", json()),
                     /* .tool_choice = */ json_value(data, "tool_choice", std::string("auto")),
diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp
@@ -17,7 +17,7 @@
 #define JSON_ASSERT GGML_ASSERT
 #include "json.hpp"
 #include "minja.hpp"
-#include "chat-handler.hpp"
+#include "chat.hpp"
 #include "chat-template.hpp"
 
 #include <random>
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
@@ -93,7 +93,7 @@ if (NOT WIN32)
     llama_target_and_test(test-grammar-parser.cpp)
     llama_target_and_test(test-grammar-integration.cpp)
     llama_target_and_test(test-llama-grammar.cpp)
-    llama_target_and_test(test-chat-handler.cpp)
+    llama_target_and_test(test-chat.cpp)
     # TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8
     if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         llama_target_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)
diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp
@@ -1,4 +1,4 @@
-#include "chat-handler.hpp"
+#include "chat.hpp"
 #include "chat-template.hpp"
 #include "llama-grammar.h"
 #include "unicode.h"
@@ -169,15 +169,15 @@ struct delta_data {
 };
 
 static delta_data init_delta(const common_chat_template & tmpl, const std::vector<std::string> & end_tokens, const json & user_message, const json & delta_message, const json & tools) {
-  common_chat_params params;
+  common_chat_inputs params;
   params.parallel_tool_calls = true;
   params.messages = json::array();
   params.messages.push_back(user_message);
   params.tools = tools;
-  auto prefix_data = common_chat_init(tmpl, params);
+  auto prefix_data = common_chat_params_init(tmpl, params);
   params.messages.push_back(delta_message);
   params.add_generation_prompt = false;
-  auto full_data = common_chat_init(tmpl, params);
+  auto full_data = common_chat_params_init(tmpl, params);
 
   std::string prefix = prefix_data.prompt;
   std::string full = full_data.prompt;
@@ -220,7 +220,7 @@ static void test_template(const common_chat_template & tmpl, const std::vector<s
   };
 
   for (const auto & tool_choice : json({"auto", "required"})) {
-    common_chat_params params;
+    common_chat_inputs params;
     params.tool_choice = tool_choice;
     params.parallel_tool_calls = true;
     params.messages = json {user_message, test_message};
@@ -301,15 +301,15 @@ static void test_template_output_parsers() {
   };
 
 
-  common_chat_params no_tools_params;
+  common_chat_inputs no_tools_params;
   no_tools_params.messages = {{{"role", "user"}, {"content", "Hey"}}};
 
-  common_chat_params tools_params = no_tools_params;
+  common_chat_inputs tools_params = no_tools_params;
   tools_params.tools = json::array();
   tools_params.tools.push_back(special_function_tool);
 
-  auto describe = [](const common_chat_template & tmpl, const common_chat_params & params) {
-    auto data = common_chat_init(tmpl, params);
+  auto describe = [](const common_chat_template & tmpl, const common_chat_inputs & params) {
+    auto data = common_chat_params_init(tmpl, params);
     return data.format;
   };
 
@@ -322,7 +322,7 @@ static void test_template_output_parsers() {
     assert_equals(std::string("generic tool calls"), describe(common_chat_template(read_file("models/templates/microsoft-Phi-3.5-mini-instruct.jinja"), "<s>", "</s>"), tools_params));
 
     // Generic tool calls doesn't generate / parse content-only messages symmetrically.
-    assert_msg_equals(msg_from_json(text_message), common_chat_init(tmpl, tools_params).parser(
+    assert_msg_equals(msg_from_json(text_message), common_chat_params_init(tmpl, tools_params).parser(
         "{\n"
         "  \"response\": \"Hello, world!\"\n"
         "}"));
