CoolFace
Modelpublic

Codeprocastinator/optimized-tinyllama-covalent

sourceHugging Faceapache-2.0updated 1y agoView on Hugging Face
0likes119downloads
chat.cpp1780 linesDownload Raw Back to common
1#include "chat.h"2#include "json-schema-to-grammar.h"3#include "log.h"4#include "minja/chat-template.hpp"5#include "minja/minja.hpp"6 7#include <optional>8 9typedef minja::chat_template common_chat_template;10 11struct common_chat_templates {12    bool has_explicit_template; // Model had builtin template or template overridde was specified.13    std::unique_ptr<common_chat_template> template_default; // always set (defaults to chatml)14    std::unique_ptr<common_chat_template> template_tool_use;15};16 17struct templates_params {18    json messages;19    json tools;20    common_chat_tool_choice tool_choice;21    json json_schema;22    bool parallel_tool_calls;23    bool stream;24    std::string grammar;25    bool add_generation_prompt = true;26    bool extract_reasoning     = true;27};28 29common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {30    if (tool_choice == "auto") {31        return COMMON_CHAT_TOOL_CHOICE_AUTO;32    }33    if (tool_choice == "none") {34        return COMMON_CHAT_TOOL_CHOICE_NONE;35    }36    if (tool_choice == "required") {37        return COMMON_CHAT_TOOL_CHOICE_REQUIRED;38    }39    throw std::runtime_error("Invalid tool_choice: " + tool_choice);40}41 42template <>43std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messages) {44    std::vector<common_chat_msg> msgs;45 46    try {47 48        if (!messages.is_array()) {49            throw std::runtime_error("Expected 'messages' to be an array, got " + messages.dump());50        }51 52        for (const auto & message : messages) {53            if (!message.is_object()) {54                throw std::runtime_error("Expected 'message' to be an object, got " + message.dump());55            }56 57            common_chat_msg msg;58            if (!message.contains("role")) {59                throw std::runtime_error("Missing 'role' in message: " + message.dump());60            }61            msg.role = message.at("role");62 63            auto has_content = message.contains("content");64            auto has_tool_calls = message.contains("tool_calls");65            if (has_content) {66                const auto & content = message.at("content");67                if (content.is_string()) {68                    msg.content = content;69                } else if (content.is_array()) {70                    for (const auto & part : content) {71                        if (!part.contains("type")) {72                            throw std::runtime_error("Missing content part type: " + part.dump());73                        }74                        const auto & type = part.at("type");75                        if (type != "text") {76                            throw std::runtime_error("Unsupported content part type: " + type.dump());77                        }78                        common_chat_msg_content_part msg_part;79                        msg_part.type = type;80                        msg_part.text = part.at("text");81                        msg.content_parts.push_back(msg_part);82                    }83                } else if (!content.is_null()) {84                    throw std::runtime_error("Invalid 'content' type: expected string or array, got " + content.dump() + " (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");85                }86            }87            if (has_tool_calls) {88                for (const auto & tool_call : message.at("tool_calls")) {89                    common_chat_tool_call tc;90                    if (!tool_call.contains("type")) {91                        throw std::runtime_error("Missing tool call type: " + tool_call.dump());92                    }93                    const auto & type = tool_call.at("type");94                    if (type != "function") {95                        throw std::runtime_error("Unsupported tool call type: " + tool_call.dump());96                    }97                    if (!tool_call.contains("function")) {98                        throw std::runtime_error("Missing tool call function: " + tool_call.dump());99                    }100                    const auto & fc = tool_call.at("function");101                    if (!fc.contains("name")) {102                        throw std::runtime_error("Missing tool call name: " + tool_call.dump());103                    }104                    tc.name = fc.at("name");105                    tc.arguments = fc.at("arguments");106                    if (tool_call.contains("id")) {107                        tc.id = tool_call.at("id");108                    }109                    msg.tool_calls.push_back(tc);110                }111            }112            if (!has_content && !has_tool_calls) {113                throw std::runtime_error("Expected 'content' or 'tool_calls' (ref: https://github.com/ggml-org/llama.cpp/issues/8367 & https://github.com/ggml-org/llama.cpp/issues/12279)");114            }115            if (message.contains("reasoning_content")) {116                msg.reasoning_content = message.at("reasoning_content");117            }118            if (message.contains("name")) {119                msg.tool_name = message.at("name");120            }121            if (message.contains("tool_call_id")) {122                msg.tool_call_id = message.at("tool_call_id");123            }124 125            msgs.push_back(msg);126        }127    } catch (const std::exception & e) {128        throw std::runtime_error("Failed to parse messages: " + std::string(e.what()) + "; messages = " + messages.dump(2));129    }130 131    return msgs;132}133 134template <>135json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {136    json messages = json::array();137    for (const auto & msg : msgs) {138        if (!msg.content.empty() && !msg.content_parts.empty()) {139            throw std::runtime_error("Cannot specify both content and content_parts");140        }141        json jmsg {142            {"role", msg.role},143        };144        if (!msg.content.empty()) {145            jmsg["content"] = msg.content;146        } else if (!msg.content_parts.empty()) {147            if (concat_typed_text) {148                std::string text;149                for (const auto & part : msg.content_parts) {150                    if (part.type != "text") {151                        LOG_WRN("Ignoring content part type: %s\n", part.type.c_str());152                        continue;153                    }154                    if (!text.empty()) {155                        text += '\n';156                    }157                    text += part.text;158                }159                jmsg["content"] = text;160            } else {161                auto & parts = jmsg["content"] = json::array();162                for (const auto & part : msg.content_parts) {163                    parts.push_back({164                        {"type", part.type},165                        {"text", part.text},166                    });167                }168            }169        } else {170            jmsg["content"] = json(); // null171        }172        if (!msg.reasoning_content.empty()) {173            jmsg["reasoning_content"] = msg.reasoning_content;174        }175        if (!msg.tool_name.empty()) {176            jmsg["name"] = msg.tool_name;177        }178        if (!msg.tool_call_id.empty()) {179            jmsg["tool_call_id"] = msg.tool_call_id;180        }181        if (!msg.tool_calls.empty()) {182            auto & tool_calls = jmsg["tool_calls"] = json::array();183            for (const auto & tool_call : msg.tool_calls) {184                json tc {185                    {"type", "function"},186                    {"function", {187                        {"name", tool_call.name},188                        {"arguments", tool_call.arguments},189                    }},190                };191                if (!tool_call.id.empty()) {192                    tc["id"] = tool_call.id;193                }194                tool_calls.push_back(tc);195            }196        }197        messages.push_back(jmsg);198    }199    return messages;200}201 202template <>203std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const std::string & messages) {204    return common_chat_msgs_parse_oaicompat(json::parse(messages));205}206 207template <>208std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {209    std::vector<common_chat_tool> result;210 211    try {212        if (!tools.is_null()) {213            if (!tools.is_array()) {214                throw std::runtime_error("Expected 'tools' to be an array, got " + tools.dump());215            }216            for (const auto & tool : tools) {217                if (!tool.contains("type")) {218                    throw std::runtime_error("Missing tool type: " + tool.dump());219                }220                const auto & type = tool.at("type");221                if (!type.is_string() || type != "function") {222                    throw std::runtime_error("Unsupported tool type: " + tool.dump());223                }224                if (!tool.contains("function")) {225                    throw std::runtime_error("Missing tool function: " + tool.dump());226                }227 228                const auto & function = tool.at("function");229                result.push_back({230                    /* .name = */ function.at("name"),231                    /* .description = */ function.at("description"),232                    /* .parameters = */ function.at("parameters").dump(),233                });234            }235        }236    } catch (const std::exception & e) {237        throw std::runtime_error("Failed to parse tools: " + std::string(e.what()) + "; tools = " + tools.dump(2));238    }239 240    return result;241}242 243template <>244std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const std::string & tools) {245    return common_chat_tools_parse_oaicompat(json::parse(tools));246}247 248template <>249json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools) {250    if (tools.empty()) {251        return json();252    }253 254    auto result = json::array();255    for (const auto & tool : tools) {256        result.push_back({257            {"type", "function"},258            {"function", {259                {"name", tool.name},260                {"description", tool.description},261                {"parameters", json::parse(tool.parameters)},262            }},263        });264    }265    return result;266}267 268bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {269    if (use_jinja) {270        try {271            common_chat_msg msg;272            msg.role = "user";273            msg.content = "test";274 275            auto tmpls = common_chat_templates_init(/* model= */ nullptr, tmpl);276 277            common_chat_templates_inputs inputs;278            inputs.messages = {msg};279 280            common_chat_templates_apply(tmpls.get(), inputs);281            return true;282        } catch (const std::exception & e) {283            LOG_ERR("%s: failed to apply template: %s\n", __func__, e.what());284            return false;285        }286    }287    llama_chat_message chat[] = {{"user", "test"}};288    const int res = llama_chat_apply_template(tmpl.c_str(), chat, 1, true, nullptr, 0);289    return res >= 0;290}291 292std::string common_chat_format_single(293        const struct common_chat_templates * tmpls,294        const std::vector<common_chat_msg> & past_msg,295        const common_chat_msg & new_msg,296        bool add_ass,297        bool use_jinja) {298 299    common_chat_templates_inputs inputs;300    inputs.use_jinja = use_jinja;301 302    std::string fmt_past_msg;303    if (!past_msg.empty()) {304        inputs.messages = past_msg;305        inputs.add_generation_prompt = false;306        fmt_past_msg = common_chat_templates_apply(tmpls, inputs).prompt;307    }308    std::ostringstream ss;309    // if the past_msg ends with a newline, we must preserve it in the formatted version310    if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {311        ss << "\n";312    };313    // format chat with new_msg314    inputs.messages.push_back(new_msg);315    inputs.add_generation_prompt = add_ass;316    auto fmt_new_msg = common_chat_templates_apply(tmpls, inputs).prompt;317    // get the diff part318    ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());319    return ss.str();320}321 322std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja) {323    common_chat_templates_inputs inputs;324    inputs.use_jinja = use_jinja;325    auto add_simple_msg = [&](auto role, auto content) {326        common_chat_msg msg;327        msg.role = role;328        msg.content = content;329        inputs.messages.push_back(msg);330    };331    add_simple_msg("system",    "You are a helpful assistant");332    add_simple_msg("user",      "Hello");333    add_simple_msg("assistant", "Hi there");334    add_simple_msg("user",      "How are you?");335    return common_chat_templates_apply(tmpls, inputs).prompt;336}337 338#define CHATML_TEMPLATE_SRC \339    "{%- for message in messages -%}\n" \340    "  {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' -}}\n" \341    "{%- endfor -%}\n" \342    "{%- if add_generation_prompt -%}\n" \343    "  {{- '<|im_start|>assistant\n' -}}\n" \344    "{%- endif -%}"345 346void common_chat_templates_free(struct common_chat_templates * tmpls) {347    delete tmpls;348}349 350bool common_chat_templates_was_explicit(const struct common_chat_templates * tmpls) {351    return tmpls->has_explicit_template;352}353 354const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant) {355    if (variant != nullptr) {356        if (strcmp(variant, "tool_use") == 0) {357            if (tmpls->template_tool_use) {358                return tmpls->template_tool_use->source().c_str();359            }360            return nullptr;361        } else {362            LOG_DBG("%s: unknown template variant: %s\n", __func__, variant);363        }364    }365    return tmpls->template_default->source().c_str();366}367 368common_chat_templates_ptr common_chat_templates_init(369    const struct llama_model * model,370    const std::string & chat_template_override,371    const std::string & bos_token_override,372    const std::string & eos_token_override)373{374    std::string default_template_src;375    std::string template_tool_use_src;376 377    bool has_explicit_template = !chat_template_override.empty();378    if (chat_template_override.empty()) {379        GGML_ASSERT(model != nullptr);380        const auto * str = llama_model_chat_template(model, /* name */ nullptr);381        if (str) {382            default_template_src = str;383            has_explicit_template = true;384        }385        str = llama_model_chat_template(model, /* name */ "tool_use");386        if (str) {387            template_tool_use_src = str;388            has_explicit_template = true;389        }390    } else {391        default_template_src = chat_template_override;392    }393    if (default_template_src.empty() || default_template_src == "chatml") {394        if (!template_tool_use_src.empty()) {395            default_template_src = template_tool_use_src;396        } else {397            default_template_src = CHATML_TEMPLATE_SRC;398        }399    }400    std::string token_bos = bos_token_override;401    std::string token_eos = eos_token_override;402    if (model) {403        const auto * vocab = llama_model_get_vocab(model);404        const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) {405            if (token == LLAMA_TOKEN_NULL) {406                if (default_template_src.find(jinja_variable_name) != std::string::npos407                    || template_tool_use_src.find(jinja_variable_name) != std::string::npos) {408                    LOG_WRN("common_chat_templates_init: warning: vocab does not have a %s token, jinja template won't work as intended.\n", name);409                }410                return std::string();411            }412            return common_token_to_piece(vocab, token, true);413        };414        token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token");415        token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token");416    }417    common_chat_templates_ptr tmpls(new common_chat_templates());418    tmpls->has_explicit_template = has_explicit_template;419    try {420        tmpls->template_default = std::make_unique<minja::chat_template>(default_template_src, token_bos, token_eos);421    } catch (const std::exception & e) {422        LOG_ERR("%s: failed to parse chat template (defaulting to chatml): %s \n", __func__, e.what());423        tmpls->template_default = std::make_unique<minja::chat_template>(CHATML_TEMPLATE_SRC, token_bos, token_eos);424    }425    if (!template_tool_use_src.empty()) {426        try {427            tmpls->template_tool_use = std::make_unique<minja::chat_template>(template_tool_use_src, token_bos, token_eos);428        } catch (const std::exception & e) {429            LOG_ERR("%s: failed to parse tool use chat template (ignoring it): %s\n", __func__, e.what());430        }431    }432    return tmpls;433}434 435std::string common_chat_format_name(common_chat_format format) {436    switch (format) {437        case COMMON_CHAT_FORMAT_CONTENT_ONLY: return "Content-only";438        case COMMON_CHAT_FORMAT_GENERIC: return "Generic";439        case COMMON_CHAT_FORMAT_MISTRAL_NEMO: return "Mistral Nemo";440        case COMMON_CHAT_FORMAT_LLAMA_3_X: return "Llama 3.x";441        case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: return "Llama 3.x with builtin tools";442        case COMMON_CHAT_FORMAT_DEEPSEEK_R1: return "DeepSeek R1";443        case COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING: return "DeepSeek R1 (extract reasoning)";444        case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: return "FireFunction v2";445        case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return "Functionary v3.2";446        case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1";447        case COMMON_CHAT_FORMAT_HERMES_2_PRO: return "Hermes 2 Pro";448        case COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING: return "Hermes 2 Pro (extract reasoning)";449        case COMMON_CHAT_FORMAT_COMMAND_R7B: return "Command R7B";450        case COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING: return "Command R7B (extract reasoning)";451        default:452            throw std::runtime_error("Unknown chat format");453    }454}455 456static bool parse_json(std::string::const_iterator & it, const std::string::const_iterator & end, json & out) {457    // // https://json.nlohmann.me/features/parsing/sax_interface/458    struct json_error_locator : public nlohmann::json_sax<json> {459        std::size_t position;460        bool found_error;461 462        json_error_locator() : position(0), found_error(false) {}463 464        bool parse_error(std::size_t position, const std::string &, const json::exception &) override { // NOLINT465            this->position = position - 1;466            this->found_error = true;467            return false;468        }469        bool null() override { return true; } // NOLINT470        bool boolean(bool) override { return true; } // NOLINT471        bool number_integer(number_integer_t) override { return true; } // NOLINT472        bool number_unsigned(number_unsigned_t) override { return true; } // NOLINT473        bool number_float(number_float_t, const string_t &) override { return true; } // NOLINT474        bool string(string_t &) override { return true; } // NOLINT475        bool binary(binary_t &) override { return true; } // NOLINT476        bool start_object(std::size_t) override { return true; } // NOLINT477        bool key(string_t &) override { return true; } // NOLINT478        bool end_object() override { return true; }479        bool start_array(std::size_t) override { return true; } // NOLINT480        bool end_array() override { return true; }481    };482    json_error_locator err_loc;483    json::sax_parse(it, end, &err_loc);484 485    std::string::const_iterator temptative_end;486    if (err_loc.found_error) {487        temptative_end = it + err_loc.position;488    } else {489        temptative_end = end;490    }491    std::string json_sub {it, temptative_end};492    try {493        out = json::parse(json_sub);494        it = temptative_end;495        return true;496    } catch (const std::exception &) {497        return false;498    }499}500 501static bool parse_literal(std::string::const_iterator & it, const std::string::const_iterator & end, const std::string & expected) {502    auto expected_it = expected.begin();503    auto tmp_it = it;504    while (tmp_it != end && expected_it != expected.end() && *tmp_it == *expected_it) {505        ++tmp_it;506        ++expected_it;507    }508    if (expected_it == expected.end()) {509        it = tmp_it;510        return true;511    }512    return false;513}514 515static std::optional<std::smatch> parse_pattern(std::string::const_iterator & it, const std::string::const_iterator & end, const std::regex & expected) {516    std::smatch match;517    if (std::regex_match(it, end, match, expected)) {518        it = match.suffix().first;519        return match;520    }521    return std::nullopt;522}523 524static void consume_spaces(std::string::const_iterator & it, const std::string::const_iterator & end) {525    while (it != end && std::isspace(*it)) {526        ++it;527    }528}529 530/**531 * Takes a prefix regex that must have 1 group to capture the function name, a closing suffix, and expects json parameters in between.532 * Aggregates the prefix, suffix and in-between text into the content.533 */534static common_chat_msg parse_json_tool_calls(535    const std::string& input,536    const std::optional<std::regex> & trigger_opt,537    const std::regex & function_regex,538    const std::regex & close_regex,539    bool allow_raw_python = false) {540    std::smatch match;541 542    common_chat_msg result;543    result.role = "assistant";544 545 546    auto end = input.end();547    auto it = input.begin();548 549    if (trigger_opt) {550        if (!std::regex_search(it, end, match, *trigger_opt)) {551            result.content = input;552            return result;553        }554        result.content = match.prefix().str();555        it = match.suffix().first;556    }557 558    while (it != end) {559        std::sregex_iterator rend;560        std::sregex_iterator rit(it, end, function_regex);561        if (rit == rend) {562            result.content += std::string(it, end);563            break;564        }565        auto name = rit->str(1);566        result.content += std::string(it, rit->prefix().second);567        it = rit->suffix().first;568 569        json arguments;570        if (parse_json(it, end, arguments)) {571            if (!std::regex_search(it, end, match, close_regex)) {572                throw std::runtime_error("Malformed input, missing closing pattern: " + input);573            }574            it = match.suffix().first;575            result.tool_calls.push_back({name, arguments.is_string() ? arguments.get<std::string>() : arguments.dump(), /* id= */ ""});576        } else {577            if (allow_raw_python && name == "python") {578                result.tool_calls.push_back({name, json({{"code", std::string(it, end)}}).dump(), /* id= */ ""});579                break;580            }581            throw std::runtime_error("Failed to parse json tool call arguments: " + input);582        }583    }584 585    if (!result.tool_calls.empty()) {586        if (!string_strip(result.content).empty()) {587            LOG_WRN("Content found with tool calls: %s\n", result.content.c_str());588        }589        result.content = "";590    }591    return result;592}593 594static common_chat_tool_call process_tool_call(const json & tool_call) {595    const auto & arguments = tool_call.at("arguments");596    return {597        /* .name = */ tool_call.at("name"),598        /* .arguments = */ arguments.is_string() ? arguments.get<std::string>() : arguments.dump(),599        /* .id = */ tool_call.contains("id") ? tool_call.at("id") : "",600    };601}602static common_chat_msg parse_prefixed_json_tool_call_array(const std::string& input, const std::string & prefix, size_t rstrip_prefix = 0) {603    auto content_end = input.find(prefix);604    size_t tc_start = std::string::npos;605 606    common_chat_msg result;607    result.role = "assistant";608    if (content_end == std::string::npos) {609        result.content = input;610    } else {611        tc_start = content_end + prefix.size() - rstrip_prefix;612        result.content = input.substr(0, content_end);613        auto tool_calls = json::parse(input.substr(tc_start));614        for (const auto & tool_call : tool_calls) {615            result.tool_calls.emplace_back(process_tool_call(tool_call));616        }617    }618    return result;619}620 621static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {622    for (const auto & tool : tools) {623        if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {624            LOG_INF("Skipping tool without function: %s", tool.dump(2).c_str());625            continue;626        }627        fn(tool);628    }629}630 631static std::string apply(632    const common_chat_template & tmpl,633    const nlohmann::ordered_json & messages,634    const nlohmann::ordered_json & tools,635    bool add_generation_prompt,636    const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())637{638    minja::chat_template_inputs tmpl_inputs;639    tmpl_inputs.messages = messages;640    tmpl_inputs.tools = tools;641    tmpl_inputs.add_generation_prompt = add_generation_prompt;642    tmpl_inputs.extra_context = extra_context;643    // TODO: add flag to control date/time, if only for testing purposes.644    // tmpl_inputs.now = std::chrono::system_clock::now();645 646    minja::chat_template_options tmpl_opts;647    // To avoid double BOS / EOS tokens, we're manually removing begining / trailing tokens648    // instead of using `chat_template_options.use_bos_token = false`, since these tokens649    // may be needed inside the template / between messages too.650    auto result = tmpl.apply(tmpl_inputs, tmpl_opts);651    if (string_starts_with(result, tmpl.bos_token())) {652        result = result.substr(tmpl.bos_token().size());653    }654    if (string_ends_with(result, tmpl.eos_token())) {655        result = result.substr(0, result.size() - tmpl.eos_token().size());656    }657    return result;658}659 660static common_chat_params common_chat_params_init_generic(const common_chat_template & tmpl, const struct templates_params & inputs) {661    common_chat_params data;662 663    auto tool_call_schemas = json::array();664    foreach_function(inputs.tools, [&](const json & tool) {665        const auto & function = tool.at("function");666        auto tool_schema = json {667            {"type", "object"},668            {"properties", {669                {"name", {670                    {"type", "string"},671                    {"const", function.at("name")},672                }},673                {"arguments", function.at("parameters")},674            }},675            {"required", json::array({"name", "arguments"})},676        };677        if (function.contains("description")) {678            tool_schema["description"] = function.at("description");679        }680        if (inputs.parallel_tool_calls) {681            tool_schema.at("properties")["id"] = {682                {"type", "string"},683                {"minLength", 4},684            };685            tool_schema.at("required").push_back("id");686        }687        tool_call_schemas.emplace_back(tool_schema);688    });689    const auto tool_call =690        inputs.parallel_tool_calls691            ? json {692                {"type", "object"},693                {"properties", {694                    {"tool_calls", {695                        {"type", "array"},696                        {"items", tool_call_schemas.size() == 1 ? tool_call_schemas[0] : json {697                            {"anyOf", tool_call_schemas},698                        }},699                        {"minItems", 1},700                    }},701                }},702                {"required", json::array({"tool_calls"})},703            }704            : json {705                {"type", "object"},706                {"properties", {707                    {"tool_call", tool_call_schemas.size() == 1 ? tool_call_schemas[0] : json {708                        {"anyOf", tool_call_schemas},709                    }},710                }},711                {"required", json::array({"tool_call"})},712            };713    const auto schema =714        inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED715            ? json {716                {"anyOf", json::array({717                    tool_call,718                    {719                        {"type", "object"},720                        {"properties", {721                            {"response", inputs.json_schema.is_null()722                                ? json {{"type", "string"}}723                                : inputs.json_schema724                            },725                        }},726                        {"required", json::array({"response"})},727                    },728                })}729            }730            : tool_call;731 732    data.grammar_lazy = false;733    data.grammar = build_grammar([&](const common_grammar_builder & builder) {734        builder.add_schema("root", schema);735    });736 737    auto tweaked_messages = common_chat_template::add_system(738        inputs.messages,739        "Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");740 741    data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);742    data.format = COMMON_CHAT_FORMAT_GENERIC;743    return data;744}745static common_chat_msg common_chat_parse_generic(const std::string & input) {746    json data = json::parse(input);747    common_chat_msg result;748    result.role = "assistant";749    if (data.contains("tool_calls")) {750        for (const auto & tool_call : data.at("tool_calls")) {751            result.tool_calls.push_back({752                tool_call.at("name"),753                tool_call.at("arguments").dump(),754                tool_call.contains("id") ? tool_call.at("id") : "",755            });756        }757    } else if (data.contains("tool_call")) {758        result.tool_calls.push_back({759            data.at("tool_call").at("name"),760            data.at("tool_call").at("arguments").dump(),761            /* id= */ "",762        });763    } else if (data.contains("response")) {764        const auto & response = data.at("response");765        result.content = response.is_string() ? response.get<std::string>() : response.dump(2);766    }767    return result;768}769 770static common_chat_params common_chat_params_init_mistral_nemo(const common_chat_template & tmpl, const struct templates_params & inputs) {771    common_chat_params data;772    data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;773    data.grammar = build_grammar([&](const common_grammar_builder & builder) {774        auto schemas = json::array();775        foreach_function(inputs.tools, [&](const json & tool) {776            const auto & function = tool.at("function");777            schemas.push_back({778                {"type", "object"},779                {"properties", {780                    // Important note: the model is probably trained to take a JSON stringified arguments value.781                    // It's hard to constrain that for now (while reusing the JSON schema conversion), so we're just expecting a plain object.782                    {"name", {783                        {"type", "string"},784                        {"const", function.at("name")},785                    }},786                    {"arguments", function.at("parameters")},787                    {"id", {788                        {"type", "string"},789                        // Nemo's template expects a 9-character alphanumeric ID.790                        {"pattern", "^[a-zA-Z0-9]{9}$"},791                    }},792                }},793                {"required", json::array({"name", "arguments", "id"})},794            });795        });796        auto schema = json {797            {"type", "array"},798            {"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},799            {"minItems", 1},800        };801        if (!inputs.parallel_tool_calls) {802            schema["maxItems"] = 1;803        }804        builder.add_rule("root", "\"[TOOL_CALLS]\" " + builder.add_schema("tool_calls", schema));805    });806    data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[TOOL_CALLS]"});807    data.preserved_tokens = {808        "[TOOL_CALLS]",809    };810    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);811    data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;812    return data;813}814static common_chat_msg common_chat_parse_mistral_nemo(const std::string & input) {815    return parse_prefixed_json_tool_call_array(input, "[TOOL_CALLS]");816}817 818static common_chat_params common_chat_params_init_command_r7b(const common_chat_template & tmpl, const struct templates_params & inputs) {819    common_chat_params data;820    data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;821    data.grammar = build_grammar([&](const common_grammar_builder & builder) {822        auto schemas = json::array();823        foreach_function(inputs.tools, [&](const json & tool) {824            const auto & function = tool.at("function");825            schemas.push_back({826                {"type", "object"},827                {"properties", {828                    {"tool_call_id", {829                        {"type", "string"},830                        // Command-R's template expects an integer string.831                        {"pattern", "^[0-9]{1,10}$"},832                    }},833                    {"tool_name", {834                        {"type", "string"},835                        {"const", function.at("name")},836                    }},837                    {"parameters", function.at("parameters")},838                }},839                {"required", json::array({"tool_call_id", "tool_name", "parameters"})},840            });841        });842        auto schema = json {843            {"type", "array"},844            {"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},845            {"minItems", 1},846        };847        if (!inputs.parallel_tool_calls) {848            schema["maxItems"] = 1;849        }850        builder.add_rule("root", "\"<|START_ACTION|>\" " + builder.add_schema("tool_calls", schema) + " \"<|END_ACTION|>\"");851    });852    data.grammar_triggers.push_back({853        COMMON_GRAMMAR_TRIGGER_TYPE_WORD,854        "<|START_ACTION|>",855    });856    data.preserved_tokens = {857        "<|START_ACTION|>",858        "<|END_ACTION|>",859        "<|START_RESPONSE|>",860        "<|END_RESPONSE|>",861        "<|START_THINKING|>",862        "<|END_THINKING|>",863    };864    auto adjusted_messages = json::array();865    for (const auto & msg : inputs.messages) {866        auto has_reasoning_content = msg.contains("reasoning_content") && msg.at("reasoning_content").is_string();867        auto has_tool_calls = msg.contains("tool_calls") && msg.at("tool_calls").is_array();868        if (has_reasoning_content && has_tool_calls) {869            auto adjusted_message = msg;870            adjusted_message["tool_plan"] = msg.at("reasoning_content");871            adjusted_message.erase("reasoning_content");872            adjusted_messages.push_back(adjusted_message);873        } else {874            adjusted_messages.push_back(msg);875        }876    }877    data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});878    data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING : COMMON_CHAT_FORMAT_COMMAND_R7B;879    return data;880}881static common_chat_msg common_chat_parse_command_r7b(const std::string & input, bool extract_reasoning) {882    static const std::regex thought_regex("(<\\|START_THINKING\\|>([\\s\\S]*?)<\\|END_THINKING\\|>)([\\s\\S]*)");883    static const std::regex action_regex("<\\|START_ACTION\\|>([\\s\\S]*?)<\\|END_ACTION\\|>");884    static const std::regex response_regex("(?:<\\|START_RESPONSE\\|>)?([\\s\\S]*?)<\\|END_RESPONSE\\|>");885 886    std::smatch match;887 888    common_chat_msg result;889    result.role = "assistant";890 891    std::string rest = input;892 893    if (std::regex_match(rest, match, thought_regex)) {894        if (extract_reasoning) {895            result.reasoning_content = match[2].str();896        } else if (!match[2].str().empty()) {897            // Let the unparsed thinking tags through in content only if their insides aren't empty.898            result.content = match[1].str();899        }900        rest = match[3].str();901    }902    if (std::regex_match(rest, match, action_regex)) {903        auto actions_str = match[1].str();904        auto actions = json::parse(actions_str);905        for (const auto & action : actions) {906            result.tool_calls.push_back({907                /* .name = */      action.at("tool_name"),908                /* .arguments = */ action.at("parameters").dump(),909                /* .id = */        action.at("tool_call_id"),910            });911        }912    } else if (std::regex_match(rest, match, response_regex)) {913        auto response = match[1].str();914        result.content += response;915    } else {916        result.content += rest;917    }918    return result;919}920 921static void expect_tool_parameters(const std::string & name, const json & parameters, const std::vector<std::string> & expected_properties) {922    if (!parameters.is_object() || !parameters.contains("type") || parameters.at("type") != "object" || !parameters.contains("properties") || !parameters.contains("required")) {923        throw std::runtime_error("Parameters of tool " + name + " must be an object w/ required properties");924    }925    const auto & parameters_properties = parameters.at("properties");926    const auto & parameters_required = parameters.at("required");927    for (const auto & prop : expected_properties) {928        if (!parameters_properties.contains(prop)) {929            throw std::runtime_error("Parameters of tool " + name + " is missing property: " + prop); // NOLINT930        }931        if (std::find(parameters_required.begin(), parameters_required.end(), json(prop)) == parameters_required.end()) {932            throw std::runtime_error("Parameters of tool " + name + " must have property marked as required: " + prop); // NOLINT933        }934    }935    if (parameters_properties.size() != expected_properties.size()) {936        throw std::runtime_error("Parameters of tool " + name + " must only have these properties:" + string_join(expected_properties, ", "));937    }938}939 940static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const common_chat_template & tmpl, const struct templates_params & inputs, bool allow_python_tag_builtin_tools) {941    auto builtin_tools = json::array();942    common_chat_params data;943    data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;944    data.grammar = build_grammar([&](const common_grammar_builder & builder) {945        std::vector<std::string> tool_rules;946 947        auto handle_builtin_tool = [&](const std::string & name, const json & parameters) {948            if (name == "wolfram_alpha" || name == "web_search" || name == "brave_search") {949                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py950                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py951                expect_tool_parameters(name, parameters, {"query"});952            } else if (name == "python" || name == "code_interpreter") {953                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py954                expect_tool_parameters(name, parameters, {"code"});955            } else {956                return false;957            }958 959            std::vector<std::string> kvs;960            for (const auto & [key, value] : parameters.at("properties").items()) {961                kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value)); // NOLINT962            }963 964            tool_rules.push_back(965                builder.add_rule(966                    name + "-call",967                    "\"<|python_tag|>" + name + ".call(\" " + string_join(kvs, " \", \" ") + " \")\""));968            builtin_tools.push_back(name);969 970            return true;971        };972 973        foreach_function(inputs.tools, [&](const json & tool) {974            const auto & function = tool.at("function");975            std::string name = function.at("name");976            auto parameters = function.at("parameters");977            builder.resolve_refs(parameters);978 979            // https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime980            if (allow_python_tag_builtin_tools) {981                handle_builtin_tool(name, parameters);982            }983            tool_rules.push_back(984                builder.add_rule(985                    name + "-call",986                    "\"{\" space "987                    "( \"\\\"type\\\"\"       space \":\" space \"\\\"function\\\"\"     space \",\" space )? "988                    "  \"\\\"name\\\"\"       space \":\" space \"\\\"" + name + "\\\"\" space \",\" space "989                    "  \"\\\"parameters\\\"\" space \":\" space " + builder.add_schema(name + "-args", parameters) + " "990                    "\"}\" space"));991        });992        // Small models may hallucinate function names so we match anything (*at the start*) that looks like the JSON of a function call, regardless of the name.993        data.grammar_triggers.push_back({994            COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,995            "\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"", // + name + "\"[\\s\\S]*",996        });997        if (!builtin_tools.empty()) {998            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});999            data.preserved_tokens.push_back("<|python_tag|>");1000        }1001        // Allow a few empty lines on top of the usual constrained json schema space rule.1002        builder.add_rule("root", string_join(tool_rules, " | "));1003    });1004    data.additional_stops.push_back("<|eom_id|>");1005    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {1006        {"tools_in_user_message", false},1007        {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},1008    });1009    data.format = allow_python_tag_builtin_tools && !builtin_tools.empty()1010        ? COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS1011        : COMMON_CHAT_FORMAT_LLAMA_3_X;1012    return data;1013}1014static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bool with_builtin_tools = false) {1015    // TODO: tighten & simplify the parser, don't accept leading text context.1016    static const std::regex function_regex(1017        "\\s*\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"([^\"]+)\"\\s*,\\s*\"parameters\"\\s*: ");1018    static const std::regex close_regex("\\}\\s*");1019    static const std::regex builtin_call_regex("<\\|python_tag\\|>\\s*([^.(]+)\\s*\\.\\s*call\\s*\\(\\s*([\\w]+)\\s*=\\s*([\\s\\S]*?)\\)");1020 1021    if (with_builtin_tools) {1022        std::smatch match;1023        if (std::regex_match(input, match, builtin_call_regex)) {1024            try {1025                auto name = match[1].str();1026                auto arg_name = match[2].str();1027                auto arg_value_str = match[3].str();1028                auto arg_value = json::parse(arg_value_str);1029 1030                common_chat_msg msg;1031                msg.role = "assistant";1032                msg.tool_calls.push_back({1033                    /* .name = */ name,1034                    /* .arguments = */ (json {1035                        {arg_name, arg_value},1036                    }).dump(),1037                    /* .id = */ "",1038                });1039                return msg;1040            } catch (const std::exception & e) {1041                LOG_WRN("Failed to parse builtin tool call arguments (%s): %s", e.what(), input.c_str());1042            }1043        }1044    }1045    return parse_json_tool_calls(input, std::nullopt, function_regex, close_regex);1046}1047 1048static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) {1049    common_chat_params data;1050    if (inputs.tools.is_array() && !inputs.tools.empty()) {1051        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED && inputs.json_schema.is_null();1052        data.grammar = build_grammar([&](const common_grammar_builder & builder) {1053            std::vector<std::string> tool_rules;1054            foreach_function(inputs.tools, [&](const json & tool) {1055                const auto & function = tool.at("function");1056                std::string name = function.at("name");1057                auto parameters = function.at("parameters");1058                builder.resolve_refs(parameters);1059                tool_rules.push_back(builder.add_rule(name + "-call",1060                    "\"<|tool▁call▁begin|>function<|tool▁sep|>" + name + "\\n"1061                    "```json\\n\" " + builder.add_schema(name + "-args", parameters) + " "1062                    "\"```<|tool▁call▁end|>\""));1063            });1064            // Distill Qwen 7B & 32B models seem confused re/ syntax of their tool call opening tag,1065            // so we accept common variants (then it's all constrained)1066            builder.add_rule("root",1067                "( \"<|tool▁calls▁begin|>\" | \"<|tool_calls_begin|>\" | \"<|tool calls begin|>\" | \"<|tool\\\\_calls\\\\_begin|>\" ) "1068                "(" + string_join(tool_rules, " | ") + ")" + (inputs.parallel_tool_calls ? "*" : "") + " "1069                "\"<|tool▁calls▁end|>\""1070                " space");1071            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|tool▁calls▁begin|>"});1072            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|tool_calls_begin|>"});1073            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|tool calls begin|>"});1074            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|tool\\_calls\\_begin|>"});1075            data.preserved_tokens = {1076                "<think>",1077                "</think>",1078                "<|tool▁calls▁begin|>",1079                "<|tool▁call▁begin|>",1080                "<|tool▁sep|>",1081                "<|tool▁call▁end|>",1082                "<|tool▁calls▁end|",1083            };1084        });1085    }1086    auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);1087 1088    // Hacks to fix the official (broken) prompt.1089    // It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,1090    // until the official template is fixed.1091    if (tmpl.source().find("{% if ns.is_tool %}{{'<|tool▁outputs▁end|>'}}") != std::string::npos) {1092        // Don't leave the chat dangling after tool results1093        if (string_ends_with(prompt, "<|tool▁outputs▁end|>")) {1094            prompt += "<|end▁of▁sentence|>";1095            if (inputs.add_generation_prompt) {1096                prompt += "<|Assistant|>";1097            }1098        }1099        // Fix up tool call delta example added by Minja1100        prompt = std::regex_replace(1101            prompt,1102            std::regex("(<|tool▁call▁end|>)[\\s\\r\\n]*(<|tool▁outputs▁begin|>|<|User|>)"),1103            "$1<|tool▁calls▁end|><|end▁of▁sentence|>$2");1104    }1105    data.prompt = prompt;1106    data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING : COMMON_CHAT_FORMAT_DEEPSEEK_R1;1107    return data;1108}1109static common_chat_msg handle_think_tag_prelude(const std::string & input, bool extract_reasoning, const std::function<common_chat_msg(const std::string &)> & rest_parser) {1110    std::smatch match;1111    static const std::regex reasoning_content_regex("((?:<think>)?([\\s\\S\\r\\n]*?)</think>)?([\\s\\S\\r\\n]*)");1112    if (std::regex_match(input, match, reasoning_content_regex)) {1113        auto rest = match[3].str();1114        auto msg = rest_parser(rest);1115        auto reasoning_content = string_strip(match[2].str());1116        if (extract_reasoning) {1117            msg.reasoning_content = reasoning_content;1118        } else if (!reasoning_content.empty()) {1119            std::ostringstream content;1120            content << "<think>" << reasoning_content << "</think>" << msg.content;1121            msg.content = content.str();1122        }1123        return msg;1124    }1125    return rest_parser(input);1126}1127static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input, bool extract_reasoning) {1128    return handle_think_tag_prelude(input, extract_reasoning, [](const std::string & input) {1129        static const std::regex function_regex("<|tool▁call▁begin|>function<|tool▁sep|>([^\n]+)\n```json\n");1130        static const std::regex close_regex("```[\\s\\r\\n]*<|tool▁call▁end|>");1131        static const 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|>");1132 1133        common_chat_msg msg;1134        msg.role = "assistant";1135        std::smatch match;1136        if (std::regex_search(input, match, tool_calls_regex)) {1137            auto tool_calls = match[1].str();1138            auto msg2 = parse_json_tool_calls(tool_calls, std::nullopt, function_regex, close_regex);1139            msg.tool_calls = std::move(msg2.tool_calls);1140        } else {1141            msg.content = input;1142        }1143        return msg;1144    });1145}1146 1147static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {1148    LOG_DBG("%s\n", __func__);1149    common_chat_params data;1150    data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {1151        {"datetime", "Jan 29 2025 13:00:00 GMT"},1152        {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},1153    });1154    if (inputs.tools.is_array() && !inputs.tools.empty()) {1155        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;1156        data.grammar = build_grammar([&](const common_grammar_builder & builder) {1157            auto schemas = json::array();1158            foreach_function(inputs.tools, [&](const json & tool) {1159                const auto & function = tool.at("function");1160                schemas.push_back({1161                    {"type", "object"},1162                    {"properties", {1163                        {"name", {1164                            {"type", "string"},1165                            {"const", function.at("name")},1166                        }},1167                        {"arguments", function.at("parameters")},1168                    }},1169                    {"required", json::array({"name", "arguments", "id"})},1170                });1171            });1172            auto schema = json {1173                {"type", "array"},1174                {"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},1175                {"minItems", 1},1176            };1177            if (!inputs.parallel_tool_calls) {1178                schema["maxItems"] = 1;1179            }1180            builder.add_rule("root", "\" functools\"? " + builder.add_schema("tool_calls", schema));1181        });1182        data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, " functools["});1183        data.preserved_tokens = {1184            " functools[",1185        };1186        data.format = COMMON_CHAT_FORMAT_FIREFUNCTION_V2;1187    } else {1188        data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;1189    }1190    return data;1191}1192static common_chat_msg common_chat_parse_firefunction_v2(const std::string & input) {1193    return parse_prefixed_json_tool_call_array(input, " functools[", /* rstrip_prefix= */ 1);1194}1195 1196static common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_template & tmpl, const struct templates_params & inputs) {1197    // >>>all\nlet's call functions>>>fn1\n{"arg1": 1...}\n>>>fn2\n{"arg1": 1...}...1198    // Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar1199    common_chat_params data;1200    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);

Showing the first 1,200 of 1780 lines. Download the file for the rest.