vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

--think to force any model to return reasoning_content (or just parse <think> for deepseek r1)

Browse source
This commit is contained in:
ochafik 2025-02-05 12:16:37 +00:00
parent 5d60cebbcc
commit 9d7c3cc51b
9 changed files with 306 additions and 145 deletions
Download patch file Download diff file Expand all files Collapse all files

9
common/arg.cpp
View file

@ -1962,6 +1962,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.use_jinja = true;
}
).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
add_opt(common_arg(
{"--think"},
"*experimental* thinking mode (default: disabled)\n"
"returns reasoning_content in messages, forcing model to think unless it supports native <think> tags (DeepSeek R1)\n"
"only supported for non-streamed responses",
[](common_params & params) {
params.think = true;
}
).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
add_opt(common_arg(
{"--chat-template"}, "JINJA_TEMPLATE",
string_format(

220
common/chat.cpp
View file

@ -12,6 +12,7 @@ std::string common_chat_format_name(common_chat_format format) {
case COMMON_CHAT_FORMAT_LLAMA_3_X: return "Llama 3.x";
case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: return "Llama 3.x with builtin tools";
case COMMON_CHAT_FORMAT_DEEPSEEK_R1: return "DeepSeek R1";
case COMMON_CHAT_FORMAT_DEEPSEEK_R1_THINK: return "DeepSeek R1 (extract <think>)";
case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: return "FireFunction v2";
case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return "Functionary v3.2";
case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1";
@ -206,6 +207,34 @@ static std::string apply(
static common_chat_params common_chat_params_init_generic(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
common_chat_params data;
json schema;
auto make_object = []() {
return json {
{"type", "object"},
{"properties", json::object()},
{"required", json::array()},
};
};
auto add_property = [](json & obj, const std::string & name, const json & schema) {
obj["properties"][name] = schema;
obj["required"].push_back(name);
};
auto add_thoughts = [&](json & obj) {
add_property(obj, "thoughts", {
{"type", "string"},
{"description", "The assistant's thoughts"},
});
};
auto make_response = [&]() {
json response_wrapper = make_object();
if (inputs.think) {
add_thoughts(response_wrapper);
}
add_property(response_wrapper, "response", inputs.json_schema.is_null() ? json {{"type", "string"}} : inputs.json_schema);
return response_wrapper;
};
std::ostringstream ss;
if (inputs.tools.is_array() && !inputs.tools.empty()) {
auto tool_call_schemas = json::array();
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
@ -232,57 +261,95 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
}
tool_call_schemas.emplace_back(tool_schema);
});
const auto tool_call =
inputs.parallel_tool_calls
? json {
{"type", "object"},
{"properties", {
{"tool_calls", {
const json tool_call = tool_call_schemas.size() == 1 ? tool_call_schemas[0] : json {{"anyOf", tool_call_schemas}};
json tool_call_wrapper = make_object();
if (inputs.think) {
add_thoughts(tool_call_wrapper);
}
if (inputs.parallel_tool_calls) {
add_property(tool_call_wrapper, "tool_calls", {
{"type", "array"},
{"items", tool_call_schemas.size() == 1 ? tool_call_schemas[0] : json {
{"anyOf", tool_call_schemas},
}},
{"items", tool_call},
{"minItems", 1},
}},
}},
{"required", json::array({"tool_calls"})},
});
} else {
add_property(tool_call_wrapper, "tool_call", tool_call);
}
: json {
{"type", "object"},
{"properties", {
{"tool_call", tool_call_schemas.size() == 1 ? tool_call_schemas[0] : json {
{"anyOf", tool_call_schemas},
}},
}},
{"required", json::array({"tool_call"})},
if (inputs.think) {
/*
This kind of turns any model into a thinking model by requiring the output to be (in TypeScript notation):
// ResponseSchema is json_schema if set, otherwisestring
Schema = ({thoughts: string} & ToolCallSchema) | {thoughts: string, response: ResponseSchema}
SchemaToolRequired = {thoughts: string} & ToolCallSchema
ToolCallSchema = SingleToolCallSchema | ParallelToolCallSchema
SingleToolCallSchema = {tool_call: ToolCall}
ParallelToolCallSchema = {tool_calls: ToolCall[]} // If parallel_tool_calls is true
ToolCall = {name: string, arguments: ParametersSchema, id?: string} // id only if parallel_tool_calls is true
ParametersSchema = tool1_params | tool2_params | ...
*/
// TODO(ochafik): make the prompts configurable (jinja template?).
ss << "You are a tool-calling assistant that thinks before it acts.\n"
"You respond in JSON format, as follows:\n"
"- First, candidly explain your thoughts about the user's request "
"and elaborate a step-by-step reasoning about your plan to satisfy it "
"(including possible tool usage / function call), pondering pros and cons, "
"widening your reasoning than narrowing down on a plan. "
"Express all of these thoughts in the `thoughts` field.\n";
}
if (inputs.tool_choice == "required") {
schema = {
{"anyOf", json::array({tool_call_wrapper, make_response()})},
};
const auto schema =
inputs.tool_choice != "required"
? json {
{"anyOf", json::array({
tool_call,
{
{"type", "object"},
{"properties", {
{"response", inputs.json_schema.is_null()
? json {{"type", "string"}}
: inputs.json_schema
},
}},
{"required", json::array({"response"})},
},
})}
if (inputs.think) {
if (inputs.parallel_tool_calls && inputs.tools.size() > 1) {
ss << "- Then if you need to perform operations or get data before responding to the user, "
"call tools by providing an array of objects with name & arguments fields in the `tool_calls` field, "
"or respond directly to the user's request in the `response` field.";
// system = "Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request";
} else {
ss << "- Then if you need to perform an operation or get data before responding to the user, "
"call a tool by providing its name & arguments in the `tool_call` field, "
"or respond directly to the user's request in the `response` field.";
}
: tool_call;
}
} else {
schema = tool_call_wrapper;
if (inputs.think) {
if (inputs.parallel_tool_calls && inputs.tools.size() > 1) {
ss << "- Then call tools by providing their names and arguments in the `tool_calls` array.";
} else {
ss << "- Then call a tool by providing its name and arguments in the `tool_call` object.";
}
}
}
ss << "- Finally, once you get results from previously requested tool calls (if you requested anys), "
"you iterate on your reasoning, update it if needed, and work towards a final response to the user's request "
"in as many iterations as needed.";
} else if (inputs.think) {
schema = make_response();
ss << "You are an assistant that thinks before it acts.\n"
"You respond in JSON format, as follows:\n"
"- First, candidly explain your thoughts about the user's request "
"and elaborate a step-by-step reasoning about your plan to satisfy it, "
"pondering pros and cons, "
"widening your reasoning than narrowing down on a plan. "
"Express all of these thoughts in the `thoughts` field.\n"
"- Then, respond directly to the user's request in the `response` field.";
}
auto system = ss.str();
data.grammar_lazy = false;
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
builder.add_schema("root", schema);
}, grammar_options);
auto tweaked_messages = common_chat_template::add_system(
inputs.messages,
"Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");
auto tweaked_messages = system.empty() ? inputs.messages : common_chat_template::add_system(inputs.messages, system);
data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.format = COMMON_CHAT_FORMAT_GENERIC;
@ -292,6 +359,9 @@ static common_chat_msg common_chat_parse_generic(const std::string & input) {
json data = json::parse(input);
common_chat_msg result;
result.role = "assistant";
if (data.contains("thoughts")) {
result.reasoning_content = data["thoughts"];
}
if (data.contains("tool_calls")) {
for (const auto & tool_call : data["tool_calls"]) {
result.tool_calls.push_back({
@ -565,7 +635,7 @@ static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bo
static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
common_chat_params data;
if (inputs.tools.is_array() && !inputs.tools.empty()) {
data.grammar_lazy = inputs.tool_choice != "required";
data.grammar_lazy = inputs.tool_choice != "required" && inputs.json_schema.is_null();
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
std::vector<std::string> tool_rules;
foreach_function(inputs.tools, [&](const json & tool) {
@ -617,27 +687,32 @@ static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_
"$1<｜tool▁calls▁end｜><｜end▁of▁sentence｜>$2");
}
data.prompt = prompt;
data.format = COMMON_CHAT_FORMAT_DEEPSEEK_R1;
data.format = inputs.think ? COMMON_CHAT_FORMAT_DEEPSEEK_R1_THINK : COMMON_CHAT_FORMAT_DEEPSEEK_R1;
return data;
}
static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input) {
static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input, bool think) {
static std::regex function_regex("<｜tool▁call▁begin｜>function<｜tool▁sep｜>([^\n]+)\n```json\n");
static std::regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>");
static std::regex reasoning_content_regex("(?:<think>([\\s\\S\\r\\n]*?)</think>)?([\\s\\S\\r\\n]*)");
static std::regex reasoning_content_regex("(<think>([\\s\\S\\r\\n]*?)</think>)?([\\s\\S\\r\\n]*)");
static std::regex tool_calls_regex("[\\s\\r\\n]*(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>)([\\s\\S\\r\\n]*?)<｜tool▁calls▁end｜>");
common_chat_msg msg;
msg.role = "assistant";
std::smatch match;
if (std::regex_match(input, match, reasoning_content_regex)) {
msg.reasoning_content = string_trim(match[1].str());
auto rest = match[2].str();
std::string rest;
if (think) {
msg.reasoning_content = string_trim(match[2].str());
} else {
msg.content = match[1].str();
}
rest = match[3].str();
if (std::regex_search(rest, match, tool_calls_regex)) {
auto tool_calls = match[1].str();
auto msg2 = parse_json_tool_calls(tool_calls, std::nullopt, function_regex, close_regex);
msg.tool_calls = std::move(msg2.tool_calls);
} else {
msg.content = std::string(rest.begin() + rest.find_first_not_of(" \r\n"), rest.end());
msg.content += std::string(rest.begin() + rest.find_first_not_of(" \r\n"), rest.end());
}
} else {
msg.content = input;
@ -953,47 +1028,66 @@ static common_chat_params common_chat_params_init_without_tools(const common_cha
}
common_chat_params common_chat_params_init(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
auto has_tools = !inputs.tools.is_null() && inputs.tool_choice != "none";
LOG_DBG("[%s] has_tools=%s\n", __func__, has_tools ? "true" : "false");
if (has_tools && !inputs.grammar.empty()) {
if (inputs.tools.is_array() && inputs.tool_choice != "none" && !inputs.grammar.empty()) {
throw std::runtime_error("Cannot specify grammar with tools");
}
const auto & src = tmpl.source();
if (src.find(">>>all") != std::string::npos) {
// Functionary prepends "all\n" to plain content outputs, so we use the parser no matter when
return common_chat_params_init_functionary_v3_2(tmpl, inputs);
}
if (src.find(" functools[") != std::string::npos) {
// Firefunction v2 requires datetime and functions in the context, even w/o tools.
return common_chat_params_init_firefunction_v2(tmpl, inputs);
}
if (src.find("<｜tool▁calls▁begin｜>") != std::string::npos) {
// DeepSeek R1: use handler in all cases except json schema (thinking / tools).
if (src.find("<｜tool▁calls▁begin｜>") != std::string::npos && inputs.json_schema.is_null()) {
return common_chat_params_init_deepseek_r1(tmpl, inputs);
}
if (!has_tools) {
// Use generic handler when forcing thoughts or JSON schema for final output
// TODO: support thinking mode and/or JSON schema in handlers below this.
if (inputs.think || inputs.json_schema.is_object()) {
return common_chat_params_init_generic(tmpl, inputs);
}
// Functionary prepends "all\n" to plain content outputs, so we use its handler in all cases.
if (src.find(">>>all") != std::string::npos) {
return common_chat_params_init_functionary_v3_2(tmpl, inputs);
}
// Firefunction v2 requires datetime and functions in the context even w/o tools, so we also use its handler in all cases.
if (src.find(" functools[") != std::string::npos) {
return common_chat_params_init_firefunction_v2(tmpl, inputs);
}
// Plain handler (no tools)
if (inputs.tools.is_null() || inputs.tool_choice == "none") {
return common_chat_params_init_without_tools(tmpl, inputs);
}
// Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
if (src.find("<tool_call>") != std::string::npos) {
return common_chat_params_init_hermes_2_pro(tmpl, inputs);
}
// Functionary v3.1 (w/ tools)
if (src.find("<|start_header_id|>") != std::string::npos
&& src.find("<function=") != std::string::npos) {
return common_chat_params_init_functionary_v3_1_llama_3_1(tmpl, inputs);
}
// Llama 3.1, 3.2, 3.3 (w/ tools)
if (src.find("<|start_header_id|>ipython<|end_header_id|>") != std::string::npos) {
auto allow_python_tag_builtin_tools = src.find("<|python_tag|>") != std::string::npos;
return common_chat_params_init_llama_3_1_tool_calls(tmpl, inputs, allow_python_tag_builtin_tools);
}
// Mistral Nemo (w/ tools)
if (src.find("[TOOL_CALLS]") != std::string::npos) {
return common_chat_params_init_mistral_nemo(tmpl, inputs);
}
// Command R7B (w/ tools)
if (src.find("<|END_THINKING|><|START_ACTION|>") != std::string::npos) {
return common_chat_params_init_command_r7b(tmpl, inputs);
}
// Generic fallback
return common_chat_params_init_generic(tmpl, inputs);
}
@ -1018,7 +1112,9 @@ common_chat_msg common_chat_parse(const std::string & input, common_chat_format
case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS:
return common_chat_parse_llama_3_1(input, /* with_builtin_tools= */ true);
case COMMON_CHAT_FORMAT_DEEPSEEK_R1:
return common_chat_parse_deepseek_r1(input);
return common_chat_parse_deepseek_r1(input, /* think= */ false);
case COMMON_CHAT_FORMAT_DEEPSEEK_R1_THINK:
return common_chat_parse_deepseek_r1(input, /* think= */ true);
case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2:
return common_chat_parse_functionary_v3_2(input);
case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1:

2
common/chat.hpp
View file

@ -19,6 +19,7 @@ struct common_chat_inputs {
bool stream;
std::string grammar;
bool add_generation_prompt = true;
bool think = false;
};
enum common_chat_format {
@ -28,6 +29,7 @@ enum common_chat_format {
COMMON_CHAT_FORMAT_LLAMA_3_X,
COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
COMMON_CHAT_FORMAT_DEEPSEEK_R1,
COMMON_CHAT_FORMAT_DEEPSEEK_R1_THINK,
COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,

1
common/common.h
View file

@ -346,6 +346,7 @@ struct common_params {
std::string chat_template = ""; // NOLINT
bool use_jinja = false; // NOLINT
bool enable_chat_template = true;
bool think = false; // return reasoning_content, force model to think unless it supports native <think> tags.
std::vector<std::string> api_keys;

4
examples/server/server.cpp
View file

@ -4052,7 +4052,7 @@ int main(int argc, char ** argv) {
}
auto body = json::parse(req.body);
json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.think, ctx_server.chat_templates);
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
@ -4065,7 +4065,7 @@ int main(int argc, char ** argv) {
// same with handle_chat_completions, but without inference part
const auto handle_apply_template = [&ctx_server, &params, &res_ok](const httplib::Request & req, httplib::Response & res) {
auto body = json::parse(req.body);
json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.think, ctx_server.chat_templates);
res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
};

24
examples/server/tests/unit/test_tool_call.py
View file

@ -439,14 +439,20 @@ def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str,
@pytest.mark.slow
@pytest.mark.parametrize("n_predict,expect_content,expect_reasoning_content,hf_repo,template_override", [
(128, "^The sum of 102 and 7 is 109.*", None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(1024, "To find the sum of.*", "I need to calculate the sum of 102 and 7.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, "To find the sum of.*", "First, I need to add the tens place.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
@pytest.mark.parametrize("n_predict,think,expect_content,expect_reasoning_content,hf_repo,template_override", [
(1024, True, "^The sum of 102 and 7 is 109.*", "^The user's request is straightforward.*", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(128, False, "^The sum of 102 and 7 is 109.*", None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(1024, True, "To find the sum of.*", "I need to calculate the sum of 102 and 7.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, False, "<think>\nI need[\\s\\S\\r\\n]*</think>\nTo find", None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, True, "To find the sum of.*", "First, I need to add the tens place.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
(1024, False, "<think>\nI need[\\s\\S\\r\\n]*</think>To find", None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
])
def test_reasoning_content(n_predict: int, expect_content: str | None, expect_reasoning_content: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
def test_thoughts(n_predict: int, think: bool, expect_content: str | None, expect_reasoning_content: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
global server
server.n_slots = 1
server.think = think
server.jinja = True
server.n_ctx = 8192 * 2
server.n_predict = n_predict
@ -470,11 +476,15 @@ def test_reasoning_content(n_predict: int, expect_content: str | None, expect_re
assert choice["message"].get("tool_calls") is None, f'Expected no tool call in {choice["message"]}'
content = choice["message"].get("content")
if expect_content is not None:
if expect_content is None:
assert content is None, f'Expected no content in {choice["message"]}'
else:
assert re.match(expect_content, content), f'Expected {expect_content}, got {content}'
reasoning_content = choice["message"].get("reasoning_content")
if expect_reasoning_content is not None:
if expect_reasoning_content is None:
assert reasoning_content is None, f'Expected no reasoning content in {choice["message"]}'
else:
assert re.match(expect_reasoning_content, reasoning_content), f'Expected {expect_reasoning_content}, got {reasoning_content}'

3
examples/server/tests/utils.py
View file

@ -78,6 +78,7 @@ class ServerProcess:
draft_max: int | None = None
no_webui: bool | None = None
jinja: bool | None = None
think: bool | None = None
chat_template: str | None = None
chat_template_file: str | None = None
@ -172,6 +173,8 @@ class ServerProcess:
server_args.append("--no-webui")
if self.jinja:
server_args.append("--jinja")
if self.think:
server_args.append("--think")
if self.chat_template:
server_args.extend(["--chat-template", self.chat_template])
if self.chat_template_file:

2
examples/server/utils.hpp
View file

@ -578,6 +578,7 @@ static json oaicompat_completion_params_parse(const json & body) {
static json oaicompat_completion_params_parse(
const json & body, /* openai api json semantics */
bool use_jinja,
bool think,
const common_chat_templates & chat_templates)
{
json llama_params;
@ -633,6 +634,7 @@ static json oaicompat_completion_params_parse(
throw std::runtime_error("Cannot use custom grammar constraints with tools.");
}
common_chat_inputs inputs;
inputs.think = think;
inputs.messages = body.at("messages");
inputs.tools = tools;
inputs.tool_choice = tool_choice;

132
tests/test-chat.cpp
View file

@ -289,11 +289,19 @@ static void test_template(const common_chat_template & tmpl, const std::vector<s
}
static void test_template_output_parsers() {
json text_message {
json message_user {
{ "role", "user" },
{ "content", "Hey there!" },
};
json message_assist {
{ "role", "assistant" },
{ "content", "Hello, world!\nWhat's up?" },
};
json text_reasoning_message {
json message_assist_thoughts_unparsed {
{ "role", "assistant" },
{ "content", "<think>I'm thinking</think>Hello, world!\nWhat's up?" },
};
json message_assist_thoughts {
{ "role", "assistant" },
{ "content", "Hello, world!\nWhat's up?" },
{ "reasoning_content", "I'm thinking" },
@ -303,7 +311,7 @@ static void test_template_output_parsers() {
{ "function", { { "name", "special_function" }, { "arguments", "{\"arg1\": 1}" } } },
}});
json tool_call_message {
json message_assist_call {
{ "role", "assistant"},
{ "content", {}},
{ "tool_calls", {
@ -316,7 +324,7 @@ static void test_template_output_parsers() {
},
}},
};
json tool_call_reasoning_message = {
json message_assist_call_thoughts = {
{ "role", "assistant" },
{ "content", nullptr },
{ "reasoning_content", "I'm\nthinking" },
@ -330,7 +338,20 @@ static void test_template_output_parsers() {
},
}},
};
json tool_call_message_with_id {
json message_assist_call_thoughts_unparsed = {
{ "role", "assistant" },
{ "content", "<think>I'm\nthinking</think>" },
{ "tool_calls", {
{
{ "type", "function" },
{ "function", {
{ "name", "special_function" },
{ "arguments", "{\"arg1\": 1}" },
}},
},
}},
};
json message_assist_call_id {
{ "role", "assistant"},
{ "content", {}},
{ "tool_calls", {
@ -347,7 +368,7 @@ static void test_template_output_parsers() {
{ "content", {} },
{ "tool_calls", tool_calls }
};
json tool_call_plan_message_with_idx {
json message_assist_call_idx {
{ "role", "assistant"},
{ "content", {}},
{ "tool_plan", "I'm not so sure"},
@ -367,7 +388,7 @@ static void test_template_output_parsers() {
{ "tool_calls", tool_calls }
};
auto python_tool_call_message = json{
auto python_message_assist_call = json{
{ "role", "assistant" },
{ "content", {} },
{ "tool_calls", json{ {
@ -382,7 +403,7 @@ static void test_template_output_parsers() {
} },
} } }
};
auto code_interpreter_tool_call_message = json{
auto code_interpreter_message_assist_call = json{
{ "role", "assistant" },
{ "content", {} },
{ "tool_calls", json{ {
@ -399,17 +420,24 @@ static void test_template_output_parsers() {
};
common_chat_inputs inputs_no_tools;
inputs_no_tools.messages = {
{ { "role", "user" }, { "content", "Hey\nThere" } }
};
inputs_no_tools.messages = json::array({message_user});
common_chat_inputs inputs_tools = inputs_no_tools;
inputs_tools.tools = json::array();
inputs_tools.tools.push_back(special_function_tool);
common_chat_inputs inputs_no_tools_think;
inputs_no_tools_think.messages = json::array({message_user});
inputs_no_tools_think.think = true;
common_chat_inputs inputs_tools_builtin = inputs_no_tools;
inputs_tools_builtin.tools = json::array();
inputs_tools_builtin.tools.push_back(python_tool);
common_chat_inputs inputs_tools;
inputs_tools.messages = json::array({message_user});
inputs_tools.tools = json::array({special_function_tool});
common_chat_inputs inputs_tools_think;
inputs_tools_think.messages = json::array({message_user});
inputs_tools_think.tools = json::array({special_function_tool});
inputs_tools_think.think = true;
common_chat_inputs inputs_tools_builtin;
inputs_tools_builtin.messages = json::array({message_user});
inputs_tools_builtin.tools = json::array({python_tool});
{
// Not supported yet
@ -423,12 +451,12 @@ static void test_template_output_parsers() {
assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_params_init(tmpl, inputs_no_tools).format);
assert_equals(COMMON_CHAT_FORMAT_COMMAND_R7B, common_chat_params_init(tmpl, inputs_tools).format);
test_template(tmpl, end_tokens, tool_call_plan_message_with_idx, tools,
test_template(tmpl, end_tokens, message_assist_call_idx, tools,
"<|START_THINKING|>I'm not so sure<|END_THINKING|>"
"<|START_ACTION|>[\n"
" {\"tool_call_id\": \"0\", \"tool_name\": \"special_function\", \"parameters\": {\"arg1\": 1}}\n"
"]<|END_ACTION|>");
test_template(tmpl, end_tokens, text_message, tools,
test_template(tmpl, end_tokens, message_assist, tools,
"<|START_RESPONSE|>Hello, world!\n"
"What's up?<|END_RESPONSE|>",
/* expect_grammar_triggered= */ false);
@ -448,12 +476,12 @@ static void test_template_output_parsers() {
// Generic tool calls doesn't generate / parse content-only messages symmetrically.
assert_msg_equals(msg_from_json(text_message),
assert_msg_equals(msg_from_json(message_assist),
common_chat_parse("{\n"
" \"response\": \"Hello, world!\\nWhat's up?\"\n"
"}",
common_chat_params_init(tmpl, inputs_tools).format));
test_template(tmpl, end_tokens, tool_call_message_with_id, tools,
test_template(tmpl, end_tokens, message_assist_call_id, tools,
"{\n"
" \"tool_calls\": [\n"
" {\n"
@ -473,9 +501,9 @@ static void test_template_output_parsers() {
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!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(
tmpl, end_tokens, tool_call_message_with_id, tools,
tmpl, end_tokens, message_assist_call_id, tools,
"[TOOL_CALLS][{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}, \"id\": \"123456789\"}]");
}
{
@ -498,12 +526,12 @@ static void test_template_output_parsers() {
inputs_tools)
.format);
test_template(tmpl, end_tokens, text_message, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, tool_call_message, tools,
test_template(tmpl, end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, message_assist_call, tools,
"<tool_call>\n"
"{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
"</tool_call>");
test_template(tmpl, end_tokens, python_tool_call_message, tools,
test_template(tmpl, end_tokens, python_message_assist_call, tools,
"<tool_call>\n"
"{\"name\": \"python\", \"arguments\": {\"code\": \"print('hey')\"}}\n"
"</tool_call>");
@ -523,12 +551,12 @@ static void test_template_output_parsers() {
inputs_tools_builtin)
.format);
// test_template(tmpl, end_tokens, text_message, tools, R"(?)", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, code_interpreter_tool_call_message, llama_3_1_tools,
// test_template(tmpl, end_tokens, message_assist, tools, R"(?)", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, code_interpreter_message_assist_call, llama_3_1_tools,
"<|python_tag|>code_interpreter.call(code=\"print('hey')\")");
test_template(tmpl, end_tokens, python_tool_call_message, tools,
test_template(tmpl, end_tokens, python_message_assist_call, tools,
"<|python_tag|>python.call(code=\"print('hey')\")");
test_template(tmpl, end_tokens, tool_call_message, tools,
test_template(tmpl, end_tokens, message_assist_call, tools,
"{\"name\": \"special_function\", \"parameters\": {\"arg1\": 1}}");
}
{
@ -538,8 +566,8 @@ static void test_template_output_parsers() {
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!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, tool_call_message, tools,
test_template(tmpl, end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, message_assist_call, tools,
"{\"name\": \"special_function\", \"parameters\": {\"arg1\": 1}}");
}
{
@ -550,8 +578,8 @@ static void test_template_output_parsers() {
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!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, tool_call_message, tools,
test_template(tmpl, end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, message_assist_call, tools,
"<function=special_function>{\"arg1\": 1}</function>");
}
{
@ -562,12 +590,12 @@ static void test_template_output_parsers() {
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, {},
test_template(tmpl, end_tokens, message_assist, {},
"all\n"
"Hello, world!\n"
"What's up?",
/* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, tool_call_message, tools,
test_template(tmpl, end_tokens, message_assist_call, tools,
"special_function\n"
"{\"arg1\": 1}");
}
@ -578,8 +606,8 @@ static void test_template_output_parsers() {
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!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, tool_call_message, tools,
test_template(tmpl, end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, message_assist_call, tools,
" functools[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]");
}
{
@ -590,10 +618,11 @@ static void test_template_output_parsers() {
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!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, text_reasoning_message, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
assert_msg_equals(msg_from_json(text_reasoning_message), common_chat_parse("<think>I'm thinking</think>Hello, world!\nWhat's up?", COMMON_CHAT_FORMAT_DEEPSEEK_R1));
// test_template(tmpl, end_tokens, tool_call_message, tools,
test_template(tmpl, end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, message_assist_thoughts, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
assert_msg_equals(msg_from_json(message_assist_thoughts_unparsed), common_chat_parse("<think>I'm thinking</think>Hello, world!\nWhat's up?", COMMON_CHAT_FORMAT_DEEPSEEK_R1));
assert_msg_equals(msg_from_json(message_assist_thoughts), common_chat_parse("<think>I'm thinking</think>Hello, world!\nWhat's up?", COMMON_CHAT_FORMAT_DEEPSEEK_R1_THINK));
// test_template(tmpl, end_tokens, message_assist_call, tools,
// "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
// "```json\n"
// "{\"arg1\": 1}\n"
@ -610,11 +639,12 @@ static void test_template_output_parsers() {
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!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, text_reasoning_message, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
assert_msg_equals(msg_from_json(text_reasoning_message), common_chat_parse("<think>I'm thinking</think>Hello, world!\nWhat's up?", COMMON_CHAT_FORMAT_DEEPSEEK_R1));
test_template(tmpl, end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
test_template(tmpl, end_tokens, message_assist_thoughts, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
assert_msg_equals(msg_from_json(message_assist_thoughts_unparsed), common_chat_parse("<think>I'm thinking</think>Hello, world!\nWhat's up?", COMMON_CHAT_FORMAT_DEEPSEEK_R1));
assert_msg_equals(msg_from_json(message_assist_thoughts), common_chat_parse("<think>I'm thinking</think>Hello, world!\nWhat's up?", COMMON_CHAT_FORMAT_DEEPSEEK_R1_THINK));
assert_msg_equals(msg_from_json(tool_call_reasoning_message),
assert_msg_equals(msg_from_json(message_assist_call_thoughts_unparsed),
common_chat_parse(
"<think>I'm\nthinking</think>\n\n"
"<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
@ -622,7 +652,15 @@ static void test_template_output_parsers() {
"{\"arg1\": 1}\n"
"```<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
COMMON_CHAT_FORMAT_DEEPSEEK_R1));
test_template(tmpl, end_tokens, tool_call_message, tools,
assert_msg_equals(msg_from_json(message_assist_call_thoughts),
common_chat_parse(
"<think>I'm\nthinking</think>\n\n"
"<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
"```json\n"
"{\"arg1\": 1}\n"
"```<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
COMMON_CHAT_FORMAT_DEEPSEEK_R1_THINK));
test_template(tmpl, end_tokens, message_assist_call, tools,
"<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
"```json\n"
"{\"arg1\": 1}\n"