Format test-chat.cpp
This commit is contained in:
Olivier Chafik
parent
5a64af6c70
commit
f223df0271
1 changed files with 340 additions and 315 deletions
|
|
@ -5,20 +5,20 @@
|
|||
//
|
||||
// cmake -B build && cmake --build build --parallel && ./build/bin/test-chat ../minja/build/tests/*.jinja 2>/dev/null
|
||||
//
|
||||
#include "chat.hpp"
|
||||
#include "chat-template.hpp"
|
||||
#include "llama-grammar.h"
|
||||
#include "unicode.h"
|
||||
|
||||
#include <fstream>
|
||||
#include <iostream>
|
||||
#include <string>
|
||||
#include <json.hpp>
|
||||
#include <string>
|
||||
|
||||
#include "chat-template.hpp"
|
||||
#include "chat.hpp"
|
||||
#include "llama-grammar.h"
|
||||
#include "unicode.h"
|
||||
|
||||
using json = nlohmann::ordered_json;
|
||||
|
||||
static common_chat_msg msg_from_json(const json & message) {
|
||||
common_chat_msg ret {
|
||||
common_chat_msg ret{
|
||||
"assistant",
|
||||
"",
|
||||
{},
|
||||
|
|
@ -40,8 +40,7 @@ static common_chat_msg msg_from_json(const json & message) {
|
|||
return ret;
|
||||
}
|
||||
|
||||
template <class T>
|
||||
static void assert_equals(const T & expected, const T & actual) {
|
||||
template <class T> static void assert_equals(const T & expected, const T & actual) {
|
||||
if (expected != actual) {
|
||||
std::cerr << "Expected: " << expected << std::endl;
|
||||
std::cerr << "Actual: " << actual << std::endl;
|
||||
|
|
@ -50,7 +49,7 @@ static void assert_equals(const T & expected, const T & actual) {
|
|||
}
|
||||
}
|
||||
|
||||
static std::string read_file(const std::string &path) {
|
||||
static std::string read_file(const std::string & path) {
|
||||
std::cerr << "# Reading: " << path << std::endl << std::flush;
|
||||
std::ifstream fs(path, std::ios_base::binary);
|
||||
if (!fs.is_open()) {
|
||||
|
|
@ -69,7 +68,8 @@ static std::string read_file(const std::string &path) {
|
|||
}
|
||||
|
||||
static std::unique_ptr<llama_grammar> build_grammar(const std::string & grammar_str) {
|
||||
return std::unique_ptr<llama_grammar>(llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root", false, nullptr, 0, nullptr, 0));
|
||||
return std::unique_ptr<llama_grammar>(
|
||||
llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root", false, nullptr, 0, nullptr, 0));
|
||||
}
|
||||
|
||||
// TODO: extract to common helper (copied from test-grammar-integration.cpp)
|
||||
|
|
@ -166,8 +166,8 @@ const auto code_interpreter_tool = json::parse(R"({
|
|||
}
|
||||
}
|
||||
})");
|
||||
const json tools = {special_function_tool, python_tool};
|
||||
const json llama_3_1_tools = {special_function_tool, code_interpreter_tool};
|
||||
const json tools = { special_function_tool, python_tool };
|
||||
const json llama_3_1_tools = { special_function_tool, code_interpreter_tool };
|
||||
|
||||
struct delta_data {
|
||||
std::string delta;
|
||||
|
|
@ -175,7 +175,9 @@ struct delta_data {
|
|||
common_chat_format format;
|
||||
};
|
||||
|
||||
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, const json & tool_choice) {
|
||||
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,
|
||||
const json & tool_choice) {
|
||||
common_chat_inputs inputs;
|
||||
inputs.parallel_tool_calls = true;
|
||||
inputs.messages = json::array();
|
||||
|
|
@ -212,7 +214,7 @@ static delta_data init_delta(const common_chat_template & tmpl, const std::vecto
|
|||
break;
|
||||
}
|
||||
}
|
||||
return {delta, params_full.grammar, params_full.format};
|
||||
return { delta, params_full.grammar, params_full.format };
|
||||
}
|
||||
|
||||
/*
|
||||
|
|
@ -220,15 +222,17 @@ static delta_data init_delta(const common_chat_template & tmpl, const std::vecto
|
|||
gets the diff, removes any end tokens and parses the result w/ the grammar, checking that
|
||||
the parsed message is the same as the test_message
|
||||
*/
|
||||
static void test_template(const common_chat_template & tmpl, const std::vector<std::string> & end_tokens, const json & test_message, const json & tools = {}, const std::string & expected_delta = "", bool skip_grammar_test = false, bool skip_parser_test = false) {
|
||||
static void test_template(const common_chat_template & tmpl, const std::vector<std::string> & end_tokens,
|
||||
const json & test_message, const json & tools = {}, const std::string & expected_delta = "",
|
||||
bool skip_grammar_test = false, bool skip_parser_test = false) {
|
||||
common_chat_msg expected_msg = msg_from_json(test_message);
|
||||
|
||||
auto user_message = json {
|
||||
{"role", "user"},
|
||||
{"content", "Hello, world!"}
|
||||
auto user_message = json{
|
||||
{ "role", "user" },
|
||||
{ "content", "Hello, world!" }
|
||||
};
|
||||
|
||||
for (const auto & tool_choice : json({"auto", "required"})) {
|
||||
for (const auto & tool_choice : json({ "auto", "required" })) {
|
||||
auto data = init_delta(tmpl, end_tokens, user_message, test_message, tools, tool_choice);
|
||||
if (!expected_delta.empty()) {
|
||||
assert_equals(expected_delta, data.delta);
|
||||
|
|
@ -250,7 +254,8 @@ static void test_template(const common_chat_template & tmpl, const std::vector<s
|
|||
// TODO: exercice lazy grammars + triggers here, instead of skipping the test
|
||||
if (!skip_grammar_test) {
|
||||
if (!match_string(data.delta, grammar.get())) {
|
||||
throw std::runtime_error("Failed to match delta against grammar:\n\n" + data.delta + "\n\nGrammar: " + data.grammar);
|
||||
throw std::runtime_error("Failed to match delta against grammar:\n\n" + data.delta +
|
||||
"\n\nGrammar: " + data.grammar);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -258,54 +263,56 @@ static void test_template(const common_chat_template & tmpl, const std::vector<s
|
|||
}
|
||||
|
||||
static void test_template_output_parsers() {
|
||||
auto text_message = json {
|
||||
{"role", "assistant"},
|
||||
{"content", "Hello, world!"},
|
||||
auto text_message = json{
|
||||
{ "role", "assistant" },
|
||||
{ "content", "Hello, world!" },
|
||||
};
|
||||
auto tool_call_message = json {
|
||||
{"role", "assistant"},
|
||||
{"content", {}},
|
||||
{"tool_calls", json {{
|
||||
{"type", "function"},
|
||||
{"function", {
|
||||
{"name", "special_function"},
|
||||
{"arguments", "{\"arg1\": 1}"}
|
||||
}},
|
||||
}}}
|
||||
auto tool_call_message = json{
|
||||
{ "role", "assistant" },
|
||||
{ "content", {} },
|
||||
{ "tool_calls", json{ {
|
||||
{ "type", "function" },
|
||||
{ "function", { { "name", "special_function" }, { "arguments", "{\"arg1\": 1}" } } },
|
||||
} } }
|
||||
};
|
||||
auto tool_call_message_with_id = json::parse(tool_call_message.dump());
|
||||
tool_call_message_with_id["tool_calls"][0]["id"] = "123456789";
|
||||
|
||||
auto python_tool_call_message = json {
|
||||
{"role", "assistant"},
|
||||
{"content", {}},
|
||||
{"tool_calls", json {{
|
||||
{"type", "function"},
|
||||
{"function", {
|
||||
{"name", "python"},
|
||||
{"arguments", {
|
||||
{"code", "print('hey')"},
|
||||
}},
|
||||
}},
|
||||
}}}
|
||||
auto python_tool_call_message = json{
|
||||
{ "role", "assistant" },
|
||||
{ "content", {} },
|
||||
{ "tool_calls", json{ {
|
||||
{ "type", "function" },
|
||||
{ "function",
|
||||
{
|
||||
{ "name", "python" },
|
||||
{ "arguments",
|
||||
{
|
||||
{ "code", "print('hey')" },
|
||||
} },
|
||||
} },
|
||||
} } }
|
||||
};
|
||||
auto code_interpreter_tool_call_message = json {
|
||||
{"role", "assistant"},
|
||||
{"content", {}},
|
||||
{"tool_calls", json {{
|
||||
{"type", "function"},
|
||||
{"function", {
|
||||
{"name", "code_interpreter"},
|
||||
{"arguments", {
|
||||
{"code", "print('hey')"},
|
||||
}},
|
||||
}},
|
||||
}}}
|
||||
auto code_interpreter_tool_call_message = json{
|
||||
{ "role", "assistant" },
|
||||
{ "content", {} },
|
||||
{ "tool_calls", json{ {
|
||||
{ "type", "function" },
|
||||
{ "function",
|
||||
{
|
||||
{ "name", "code_interpreter" },
|
||||
{ "arguments",
|
||||
{
|
||||
{ "code", "print('hey')" },
|
||||
} },
|
||||
} },
|
||||
} } }
|
||||
};
|
||||
|
||||
|
||||
common_chat_inputs inputs_no_tools;
|
||||
inputs_no_tools.messages = {{{"role", "user"}, {"content", "Hey"}}};
|
||||
inputs_no_tools.messages = {
|
||||
{ { "role", "user" }, { "content", "Hey" } }
|
||||
};
|
||||
|
||||
common_chat_inputs inputs_tools = inputs_no_tools;
|
||||
inputs_tools.tools = json::array();
|
||||
|
|
@ -315,147 +322,163 @@ static void test_template_output_parsers() {
|
|||
inputs_tools_builtin.tools = json::array();
|
||||
inputs_tools_builtin.tools.push_back(python_tool);
|
||||
|
||||
// {
|
||||
// const common_chat_template tmpl(read_file(
|
||||
// "models/templates/google-gemma-2-2b-it.jinja"), "<s>", "</s>");
|
||||
// std::vector<std::string> end_tokens { "<end_of_turn>" };
|
||||
|
||||
// assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_params_init(tmpl, inputs_no_tools).format);
|
||||
// assert_equals(COMMON_CHAT_FORMAT_GENERIC, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
// assert_equals(COMMON_CHAT_FORMAT_GENERIC, common_chat_params_init(common_chat_template(read_file(
|
||||
// "models/templates/microsoft-Phi-3.5-mini-instruct.jinja"), "<s>", "</s>"), inputs_tools).format);
|
||||
|
||||
// // Generic tool calls doesn't generate / parse content-only messages symmetrically.
|
||||
|
||||
// assert_msg_equals(msg_from_json(text_message), common_chat_parse(
|
||||
// "{\n"
|
||||
// " \"response\": \"Hello, world!\"\n"
|
||||
// "}",
|
||||
// common_chat_params_init(tmpl, inputs_tools).format));
|
||||
// test_template(tmpl, end_tokens, tool_call_message_with_id, tools,
|
||||
// "{\n"
|
||||
// " \"tool_calls\": [\n"
|
||||
// " {\n"
|
||||
// " \"name\": \"special_function\",\n"
|
||||
// " \"arguments\": {\n"
|
||||
// " \"arg1\": 1\n"
|
||||
// " },\n"
|
||||
// " \"id\": \"123456789\"\n"
|
||||
// " }\n"
|
||||
// " ]\n"
|
||||
// "}");
|
||||
// }
|
||||
// {
|
||||
// const common_chat_template tmpl(read_file(
|
||||
// "models/templates/mistralai-Mistral-Nemo-Instruct-2407.jinja"), "<s>", "</s>");
|
||||
// std::vector<std::string> end_tokens { "</s>" };
|
||||
|
||||
// assert_equals(COMMON_CHAT_FORMAT_MISTRAL_NEMO, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
|
||||
// test_template(tmpl, end_tokens, text_message, tools, "Hello, world!", /* skip_grammar_test= */ true);
|
||||
// test_template(tmpl, end_tokens, tool_call_message_with_id, tools,
|
||||
// "[TOOL_CALLS][{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}, \"id\": \"123456789\"}]",
|
||||
// /* skip_grammar_test= */ true);
|
||||
// }
|
||||
// {
|
||||
// const common_chat_template tmpl(read_file(
|
||||
// "models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja"), "<s>", "</s>");
|
||||
// std::vector<std::string> end_tokens { "<|im_end|>" };
|
||||
|
||||
// assert_equals(COMMON_CHAT_FORMAT_HERMES_2_PRO, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
// assert_equals(COMMON_CHAT_FORMAT_HERMES_2_PRO, common_chat_params_init(common_chat_template(read_file(
|
||||
// "models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja"), "<s>", "</s>"), inputs_tools).format);
|
||||
// assert_equals(COMMON_CHAT_FORMAT_HERMES_2_PRO, common_chat_params_init(common_chat_template(read_file(
|
||||
// "models/templates/Qwen-Qwen2.5-7B-Instruct.jinja"), "<s>", "</s>"), inputs_tools).format);
|
||||
|
||||
// test_template(tmpl, end_tokens, text_message, tools, "Hello, world!", /* skip_grammar_test= */ true);
|
||||
// test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
// "<tool_call>\n"
|
||||
// "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
|
||||
// "</tool_call>");
|
||||
// test_template(tmpl, end_tokens, python_tool_call_message, tools,
|
||||
// "<tool_call>\n"
|
||||
// "{\"name\": \"python\", \"arguments\": {\"code\": \"print('hey')\"}}\n"
|
||||
// "</tool_call>");
|
||||
// }
|
||||
// {
|
||||
// const common_chat_template tmpl(read_file(
|
||||
// "models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja"), "<s>", "</s>");
|
||||
// std::vector<std::string> end_tokens { "<|eom_id|>", "<|eot_id|>" };
|
||||
|
||||
// assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
// assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS, common_chat_params_init(tmpl, inputs_tools_builtin).format);
|
||||
// assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS, common_chat_params_init(common_chat_template(read_file(
|
||||
// "models/templates/meta-llama-Llama-3.3-70B-Instruct.jinja"), "<s>", "</s>"), inputs_tools_builtin).format);
|
||||
|
||||
// // test_template(tmpl, end_tokens, text_message, tools, R"(?)", /* skip_grammar_test= */ true);
|
||||
// test_template(tmpl, end_tokens, code_interpreter_tool_call_message, llama_3_1_tools,
|
||||
// "<|python_tag|>code_interpreter.call(code=\"print('hey')\")");
|
||||
// test_template(tmpl, end_tokens, python_tool_call_message, tools,
|
||||
// "<|python_tag|>python.call(code=\"print('hey')\")");
|
||||
// test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
// "{\"name\": \"special_function\", \"parameters\": {\"arg1\": 1}}");
|
||||
// }
|
||||
// {
|
||||
// const common_chat_template tmpl(read_file(
|
||||
// "models/templates/meta-llama-Llama-3.2-3B-Instruct.jinja"), "<s>", "</s>");
|
||||
// std::vector<std::string> end_tokens { "<|eom_id|>", "<|eot_id|>" };
|
||||
|
||||
// assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
|
||||
// test_template(tmpl, end_tokens, text_message, tools,
|
||||
// "Hello, world!", /* skip_grammar_test= */ true);
|
||||
// test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
// "{\"name\": \"special_function\", \"parameters\": {\"arg1\": 1}}");
|
||||
// }
|
||||
// {
|
||||
// const common_chat_template tmpl(read_file(
|
||||
// "models/templates/meetkai-functionary-medium-v3.1.jinja"), "<s>", "</s>");
|
||||
// std::vector<std::string> end_tokens { "<|eom_id|>", "<|eot_id|>" };
|
||||
|
||||
// assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
|
||||
// test_template(tmpl, end_tokens, text_message, tools,
|
||||
// "Hello, world!", /* skip_grammar_test= */ true);
|
||||
// test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
// "<function=special_function>{\"arg1\": 1}</function>");
|
||||
// }
|
||||
{
|
||||
const common_chat_template tmpl(read_file(
|
||||
"models/templates/meetkai-functionary-medium-v3.2.jinja"), "<s>", "</s>");
|
||||
std::vector<std::string> end_tokens { "<|eom_id|>", "<|eot_id|>" };
|
||||
const common_chat_template tmpl(read_file("models/templates/google-gemma-2-2b-it.jinja"), "<s>", "</s>");
|
||||
std::vector<std::string> end_tokens{ "<end_of_turn>" };
|
||||
|
||||
assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_params_init(tmpl, inputs_no_tools).format);
|
||||
assert_equals(COMMON_CHAT_FORMAT_GENERIC, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
assert_equals(COMMON_CHAT_FORMAT_GENERIC,
|
||||
common_chat_params_init(
|
||||
common_chat_template(read_file("models/templates/microsoft-Phi-3.5-mini-instruct.jinja"),
|
||||
"<s>", "</s>"),
|
||||
inputs_tools)
|
||||
.format);
|
||||
|
||||
// Generic tool calls doesn't generate / parse content-only messages symmetrically.
|
||||
|
||||
assert_msg_equals(msg_from_json(text_message),
|
||||
common_chat_parse("{\n"
|
||||
" \"response\": \"Hello, world!\"\n"
|
||||
"}",
|
||||
common_chat_params_init(tmpl, inputs_tools).format));
|
||||
test_template(tmpl, end_tokens, tool_call_message_with_id, tools,
|
||||
"{\n"
|
||||
" \"tool_calls\": [\n"
|
||||
" {\n"
|
||||
" \"name\": \"special_function\",\n"
|
||||
" \"arguments\": {\n"
|
||||
" \"arg1\": 1\n"
|
||||
" },\n"
|
||||
" \"id\": \"123456789\"\n"
|
||||
" }\n"
|
||||
" ]\n"
|
||||
"}");
|
||||
}
|
||||
{
|
||||
const common_chat_template tmpl(read_file("models/templates/mistralai-Mistral-Nemo-Instruct-2407.jinja"), "<s>",
|
||||
"</s>");
|
||||
std::vector<std::string> end_tokens{ "</s>" };
|
||||
|
||||
assert_equals(COMMON_CHAT_FORMAT_MISTRAL_NEMO, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
|
||||
test_template(tmpl, end_tokens, text_message, tools, "Hello, world!", /* skip_grammar_test= */ true);
|
||||
test_template(
|
||||
tmpl, end_tokens, tool_call_message_with_id, tools,
|
||||
"[TOOL_CALLS][{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}, \"id\": \"123456789\"}]",
|
||||
/* skip_grammar_test= */ true);
|
||||
}
|
||||
{
|
||||
const common_chat_template tmpl(
|
||||
read_file("models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja"), "<s>", "</s>");
|
||||
std::vector<std::string> end_tokens{ "<|im_end|>" };
|
||||
|
||||
assert_equals(COMMON_CHAT_FORMAT_HERMES_2_PRO, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
assert_equals(
|
||||
COMMON_CHAT_FORMAT_HERMES_2_PRO,
|
||||
common_chat_params_init(
|
||||
common_chat_template(read_file("models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja"),
|
||||
"<s>", "</s>"),
|
||||
inputs_tools)
|
||||
.format);
|
||||
assert_equals(
|
||||
COMMON_CHAT_FORMAT_HERMES_2_PRO,
|
||||
common_chat_params_init(
|
||||
common_chat_template(read_file("models/templates/Qwen-Qwen2.5-7B-Instruct.jinja"), "<s>", "</s>"),
|
||||
inputs_tools)
|
||||
.format);
|
||||
|
||||
test_template(tmpl, end_tokens, text_message, tools, "Hello, world!", /* skip_grammar_test= */ true);
|
||||
test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
"<tool_call>\n"
|
||||
"{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
|
||||
"</tool_call>");
|
||||
test_template(tmpl, end_tokens, python_tool_call_message, tools,
|
||||
"<tool_call>\n"
|
||||
"{\"name\": \"python\", \"arguments\": {\"code\": \"print('hey')\"}}\n"
|
||||
"</tool_call>");
|
||||
}
|
||||
{
|
||||
const common_chat_template tmpl(read_file("models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja"), "<s>",
|
||||
"</s>");
|
||||
std::vector<std::string> end_tokens{ "<|eom_id|>", "<|eot_id|>" };
|
||||
|
||||
assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
|
||||
common_chat_params_init(tmpl, inputs_tools_builtin).format);
|
||||
assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
|
||||
common_chat_params_init(
|
||||
common_chat_template(read_file("models/templates/meta-llama-Llama-3.3-70B-Instruct.jinja"),
|
||||
"<s>", "</s>"),
|
||||
inputs_tools_builtin)
|
||||
.format);
|
||||
|
||||
// test_template(tmpl, end_tokens, text_message, tools, R"(?)", /* skip_grammar_test= */ true);
|
||||
test_template(tmpl, end_tokens, code_interpreter_tool_call_message, llama_3_1_tools,
|
||||
"<|python_tag|>code_interpreter.call(code=\"print('hey')\")");
|
||||
test_template(tmpl, end_tokens, python_tool_call_message, tools,
|
||||
"<|python_tag|>python.call(code=\"print('hey')\")");
|
||||
test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
"{\"name\": \"special_function\", \"parameters\": {\"arg1\": 1}}");
|
||||
}
|
||||
{
|
||||
const common_chat_template tmpl(read_file("models/templates/meta-llama-Llama-3.2-3B-Instruct.jinja"), "<s>",
|
||||
"</s>");
|
||||
std::vector<std::string> end_tokens{ "<|eom_id|>", "<|eot_id|>" };
|
||||
|
||||
assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
|
||||
test_template(tmpl, end_tokens, text_message, tools, "Hello, world!", /* skip_grammar_test= */ true);
|
||||
test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
"{\"name\": \"special_function\", \"parameters\": {\"arg1\": 1}}");
|
||||
}
|
||||
{
|
||||
const common_chat_template tmpl(read_file("models/templates/meetkai-functionary-medium-v3.1.jinja"), "<s>",
|
||||
"</s>");
|
||||
std::vector<std::string> end_tokens{ "<|eom_id|>", "<|eot_id|>" };
|
||||
|
||||
assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
|
||||
common_chat_params_init(tmpl, inputs_tools).format);
|
||||
|
||||
test_template(tmpl, end_tokens, text_message, tools, "Hello, world!", /* skip_grammar_test= */ true);
|
||||
test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
"<function=special_function>{\"arg1\": 1}</function>");
|
||||
}
|
||||
{
|
||||
const common_chat_template tmpl(read_file("models/templates/meetkai-functionary-medium-v3.2.jinja"), "<s>",
|
||||
"</s>");
|
||||
std::vector<std::string> end_tokens{ "<|eom_id|>", "<|eot_id|>" };
|
||||
|
||||
assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2, common_chat_params_init(tmpl, inputs_no_tools).format);
|
||||
assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
|
||||
test_template(tmpl, end_tokens, text_message, {},
|
||||
"all\n"
|
||||
"Hello, world!", /* skip_grammar_test= */ true);
|
||||
"Hello, world!",
|
||||
/* skip_grammar_test= */ true);
|
||||
test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
"special_function\n"
|
||||
"{\"arg1\": 1}");
|
||||
}
|
||||
{
|
||||
const common_chat_template tmpl(read_file(
|
||||
"models/templates/fireworks-ai-llama-3-firefunction-v2.jinja"), "<s>", "</s>");
|
||||
std::vector<std::string> end_tokens { "<|eot_id|>" };
|
||||
const common_chat_template tmpl(read_file("models/templates/fireworks-ai-llama-3-firefunction-v2.jinja"), "<s>",
|
||||
"</s>");
|
||||
std::vector<std::string> end_tokens{ "<|eot_id|>" };
|
||||
|
||||
assert_equals(COMMON_CHAT_FORMAT_FIREFUNCTION_V2, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
|
||||
test_template(tmpl, end_tokens, text_message, tools,
|
||||
"Hello, world!", /* skip_grammar_test= */ true);
|
||||
test_template(tmpl, end_tokens, text_message, tools, "Hello, world!", /* skip_grammar_test= */ true);
|
||||
test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
" functools[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]");
|
||||
}
|
||||
{
|
||||
const common_chat_template tmpl(read_file(
|
||||
"models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja"), "<s>", "</s>");
|
||||
std::vector<std::string> end_tokens { "<｜end▁of▁sentence｜>" };
|
||||
const common_chat_template tmpl(read_file("models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja"),
|
||||
"<s>", "</s>");
|
||||
std::vector<std::string> end_tokens{ "<｜end▁of▁sentence｜>" };
|
||||
|
||||
assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1, common_chat_params_init(tmpl, inputs_tools).format);
|
||||
|
||||
test_template(tmpl, end_tokens, text_message, tools,
|
||||
"Hello, world!", /* skip_grammar_test= */ true);
|
||||
test_template(tmpl, end_tokens, text_message, tools, "Hello, world!", /* skip_grammar_test= */ true);
|
||||
test_template(tmpl, end_tokens, tool_call_message, tools,
|
||||
"<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
|
||||
"```json\n"
|
||||
|
|
@ -464,12 +487,14 @@ static void test_template_output_parsers() {
|
|||
}
|
||||
}
|
||||
|
||||
int main(int argc, char **argv) {
|
||||
int main(int argc, char ** argv) {
|
||||
#ifndef _WIN32
|
||||
if (argc > 1) {
|
||||
common_chat_inputs inputs;
|
||||
inputs.messages = {{{"role", "user"}, {"content", "Hey"}}};
|
||||
inputs.tools = json::array({special_function_tool});
|
||||
inputs.messages = {
|
||||
{ { "role", "user" }, { "content", "Hey" } }
|
||||
};
|
||||
inputs.tools = json::array({ special_function_tool });
|
||||
|
||||
std::cout << "| Template | Format |\n";
|
||||
std::cout << "|----------|--------|\n";
|
||||
|
|
@ -483,10 +508,10 @@ int main(int argc, char **argv) {
|
|||
common_chat_template tmpl(read_file(path), "", "");
|
||||
auto parts = string_split(path, "/");
|
||||
auto name = parts[parts.size() - 1];
|
||||
std::cout << "| " << name << " | " << common_chat_format_name(common_chat_params_init(tmpl, inputs).format) << " |\n";
|
||||
std::cout << "| " << name << " | " << common_chat_format_name(common_chat_params_init(tmpl, inputs).format)
|
||||
<< " |\n";
|
||||
}
|
||||
}
|
||||
else
|
||||
} else
|
||||
#endif
|
||||
{
|
||||
test_template_output_parsers();
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue