tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-07 05:23:38 +00:00
Code Issues Packages Projects Releases Wiki Activity

server : support jinja extra template kwargs (Qwen3 enable_thinking feature), from command line and from client (#13196)

Browse Source 
* initial commit for handling extra template kwargs

* enable_thinking and assistant prefill cannot be enabled at the same time

* can set chat_template_kwargs in command line

* added doc

* fixed formatting

* add support for extra context in generic template init

* coding standard: common/chat.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* coding standard:  common/chat.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Apply suggestions from code review

coding standard: cosmetic changes

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* fix merge conflict

* chat.cpp: simplify calls to apply to ensure systematic propagation of extra_context (+ the odd existing additional_context)

* normalize environment variable name

* simplify code

* prefill cannot be used with thinking models

* compatibility with the new reasoning-budget parameter

* fix prefill for non thinking models

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Olivier Chafik <olivier.chafik@gmail.com>
This commit is contained in:
matteo
2025-06-29 20:02:53 +02:00
committed by GitHub
parent 83790b0e7e
commit caf5681fcb
7 changed files with 69 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
common/chat.cpp
View File

@ -17,6 +17,8 @@
#include <string>
#include <vector>
using json = nlohmann::ordered_json;
static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
auto time = std::chrono::system_clock::to_time_t(now);
auto local_time = *std::localtime(&time);
@ -140,6 +142,7 @@ struct templates_params {
bool add_generation_prompt = true;
bool enable_thinking = true;
std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
json extra_context;
};
common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@ -720,16 +723,23 @@ static void foreach_function(const json & tools, const std::function<void(const
static std::string apply(
const common_chat_template & tmpl,
const nlohmann::ordered_json & messages,
const nlohmann::ordered_json & tools,
bool add_generation_prompt,
const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())
const struct templates_params & inputs,
const std::optional<json> & messages_override = std::nullopt,
const std::optional<json> & tools_override = std::nullopt,
const std::optional<json> & additional_context = std::nullopt)
{
minja::chat_template_inputs tmpl_inputs;
tmpl_inputs.messages = messages;
tmpl_inputs.tools = tools;
tmpl_inputs.add_generation_prompt = add_generation_prompt;
tmpl_inputs.extra_context = extra_context;
tmpl_inputs.messages = messages_override ? *messages_override : inputs.messages;
if (tools_override) {
tmpl_inputs.tools = *tools_override;
} else {
tmpl_inputs.tools = inputs.tools.empty() ? json() : inputs.tools;
}
tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
tmpl_inputs.extra_context = inputs.extra_context;
if (additional_context) {
tmpl_inputs.extra_context.merge_patch(*additional_context);
}
// TODO: add flag to control date/time, if only for testing purposes.
// tmpl_inputs.now = std::chrono::system_clock::now();
@ -828,7 +838,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
inputs.messages,
"Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");
data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
data.format = COMMON_CHAT_FORMAT_GENERIC;
return data;
}
@ -904,7 +914,7 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
data.preserved_tokens = {
"[TOOL_CALLS]",
};
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs);
data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
return data;
}
@ -934,7 +944,7 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
adjusted_messages.push_back(msg);
}
}
data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
data.prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
if (string_ends_with(data.prompt, "<|START_THINKING|>")) {
if (!inputs.enable_thinking) {
@ -1122,7 +1132,7 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
} else {
data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
}
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, json {
{"date_string", format_time(inputs.now, "%d %b %Y")},
{"tools_in_user_message", false},
{"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
@ -1187,7 +1197,7 @@ static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool w
static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
auto prompt = apply(tmpl, inputs);
// Hacks to fix the official (broken) prompt.
// It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,
@ -1282,7 +1292,7 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
LOG_DBG("%s\n", __func__);
common_chat_params data;
data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
{"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
{"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
});
@ -1338,7 +1348,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
// Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar
// If the function is python, we also allow raw python code (if the line after `python\n` doesn't start w/ opening `{`), which the model seems to prefer for multiline code.
common_chat_params data;
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs);
data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
if (inputs.tools.is_array() && !inputs.tools.empty()) {
data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@ -1465,7 +1475,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
}
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs);
// TODO: if (has_raw_python)
return data;
}
@ -1498,14 +1508,15 @@ static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser
static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
json additional_context = {
json extra_context = json {
{"enable_thinking", inputs.enable_thinking},
};
extra_context.update(inputs.extra_context);
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, additional_context);
data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, extra_context);
data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
if (string_ends_with(data.prompt, "<think>\n")) {
if (!inputs.enable_thinking) {
if (!extra_context["enable_thinking"]) {
data.prompt += "</think>";
} else {
data.thinking_forced_open = true;
@ -1691,7 +1702,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs);
data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
data.grammar_lazy = false;
if (!inputs.json_schema.is_null()) {
@ -1722,6 +1733,12 @@ static common_chat_params common_chat_templates_apply_jinja(
params.enable_thinking = inputs.enable_thinking;
params.grammar = inputs.grammar;
params.now = inputs.now;
params.extra_context = json::object();
for (auto el : inputs.chat_template_kwargs) {
params.extra_context[el.first] = json::parse(el.second);
}
if (!inputs.json_schema.empty()) {
params.json_schema = json::parse(inputs.json_schema);
}