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

268
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";
@ -206,83 +207,149 @@ 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;
auto tool_call_schemas = json::array();
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
auto tool_schema = json {
json schema;
auto make_object = []() {
return json {
{"type", "object"},
{"properties", {
{"name", {
{"type", "string"},
{"const", function["name"]},
}},
{"arguments", function["parameters"]},
}},
{"required", json::array({"name", "arguments"})},
{"properties", json::object()},
{"required", json::array()},
};
if (function.contains("description")) {
tool_schema["description"] = function["description"];
};
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"];
auto tool_schema = json {
{"type", "object"},
{"properties", {
{"name", {
{"type", "string"},
{"const", function["name"]},
}},
{"arguments", function["parameters"]},
}},
{"required", json::array({"name", "arguments"})},
};
if (function.contains("description")) {
tool_schema["description"] = function["description"];
}
if (inputs.parallel_tool_calls) {
tool_schema["properties"]["id"] = {
{"type", "string"},
{"minLength", 4},
};
tool_schema["required"].push_back("id");
}
tool_call_schemas.emplace_back(tool_schema);
});
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) {
tool_schema["properties"]["id"] = {
{"type", "string"},
{"minLength", 4},
};
tool_schema["required"].push_back("id");
add_property(tool_call_wrapper, "tool_calls", {
{"type", "array"},
{"items", tool_call},
{"minItems", 1},
});
} else {
add_property(tool_call_wrapper, "tool_call", tool_call);
}
tool_call_schemas.emplace_back(tool_schema);
});
const auto tool_call =
inputs.parallel_tool_calls
? json {
{"type", "object"},
{"properties", {
{"tool_calls", {
{"type", "array"},
{"items", tool_call_schemas.size() == 1 ? tool_call_schemas[0] : json {
{"anyOf", tool_call_schemas},
}},
{"minItems", 1},
}},
}},
{"required", json::array({"tool_calls"})},
}
: 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: