llama.cpp/tools/server/utils.hpp

#pragma once

#include "common.h"
#include "log.h"
#include "llama.h"
#include "arg.h" // common_remote_get_content
#include "base64.hpp"
#include "mtmd.h"

// increase max payload length to allow use of larger context size
#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
// disable Nagle's algorithm
#define CPPHTTPLIB_TCP_NODELAY true
#include "httplib.h"

// Change JSON_ASSERT from assert() to GGML_ASSERT:
#define JSON_ASSERT GGML_ASSERT
#include "json.hpp"
#include "chat.h"

#include <random>
#include <sstream>
#include <string>
#include <vector>
#include <memory>
#include <cinttypes>

#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo"

using json = nlohmann::ordered_json;

#define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SLT_WRN(slot, fmt, ...) LOG_WRN("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SLT_ERR(slot, fmt, ...) LOG_ERR("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SLT_DBG(slot, fmt, ...) LOG_DBG("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)

#define SRV_INF(fmt, ...) LOG_INF("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define SRV_WRN(fmt, ...) LOG_WRN("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define SRV_ERR(fmt, ...) LOG_ERR("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define SRV_DBG(fmt, ...) LOG_DBG("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)

#define QUE_INF(fmt, ...) LOG_INF("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_WRN(fmt, ...) LOG_WRN("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_ERR(fmt, ...) LOG_ERR("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_DBG(fmt, ...) LOG_DBG("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)

using raw_buffer = std::vector<uint8_t>;

template <typename T>
static T json_value(const json & body, const std::string & key, const T & default_value) {
    // Fallback null to default value
    if (body.contains(key) && !body.at(key).is_null()) {
        try {
            return body.at(key);
        } catch (NLOHMANN_JSON_NAMESPACE::detail::type_error const &) {
            LOG_WRN("Wrong type supplied for parameter '%s'. Expected '%s', using default value\n", key.c_str(), json(default_value).type_name());
            return default_value;
        }
    } else {
        return default_value;
    }
}

const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);

// thin wrapper around common_grammar_trigger with (de)serialization functions
struct server_grammar_trigger {
    common_grammar_trigger value;

    server_grammar_trigger() = default;
    server_grammar_trigger(const common_grammar_trigger & value) : value(value) {}
    server_grammar_trigger(const json & in) {
        value.type = (common_grammar_trigger_type) in.at("type").get<int>();
        value.value = in.at("value").get<std::string>();
        if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
            value.token = (llama_token) in.at("token").get<int>();
        }
    }

    json to_json() const {
        json out {
            {"type", (int) value.type},
            {"value", value.value},
        };
        if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
            out["token"] = (int) value.token;
        }
        return out;
    }
};

//
// tokenizer and input processing utils
//

static bool json_is_array_of_numbers(const json & data) {
    if (data.is_array()) {
        for (const auto & e : data) {
            if (!e.is_number_integer()) {
                return false;
            }
        }
        return true;
    }
    return false;
}

// is array having BOTH numbers & strings?
static bool json_is_array_of_mixed_numbers_strings(const json & data) {
    bool seen_string = false;
    bool seen_number = false;
    if (data.is_array()) {
        for (const auto & e : data) {
            seen_string |= e.is_string();
            seen_number |= e.is_number_integer();
            if (seen_number && seen_string) {
                return true;
            }
        }
    }
    return false;
}

// get value by path(key1 / key2)
static json json_get_nested_values(const std::vector<std::string> & paths, const json & js) {
    json result = json::object();

    for (const std::string & path : paths) {
        json current = js;
        const auto keys = string_split<std::string>(path, /*separator*/ '/');
        bool valid_path = true;
        for (const std::string & k : keys) {
            if (valid_path && current.is_object() && current.contains(k)) {
                current = current[k];
            } else {
                valid_path = false;
            }
        }
        if (valid_path) {
            result[path] = current;
        }
    }
    return result;
}

/**
 * this handles 2 cases:
 * - only string, example: "string"
 * - mixed string and tokens, example: [12, 34, "string", 56, 78]
 */
static llama_tokens tokenize_mixed(const llama_vocab * vocab, const json & json_prompt, bool add_special, bool parse_special) {
    // If `add_bos` is true, we only add BOS, when json_prompt is a string,
    // or the first element of the json_prompt array is a string.
    llama_tokens prompt_tokens;

    if (json_prompt.is_array()) {
        bool first = true;
        for (const auto & p : json_prompt) {
            if (p.is_string()) {
                auto s = p.template get<std::string>();

                llama_tokens p;
                if (first) {
                    p = common_tokenize(vocab, s, add_special, parse_special);
                    first = false;
                } else {
                    p = common_tokenize(vocab, s, false, parse_special);
                }

                prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
            } else {
                if (first) {
                    first = false;
                }

                prompt_tokens.push_back(p.template get<llama_token>());
            }
        }
    } else {
        auto s = json_prompt.template get<std::string>();
        prompt_tokens = common_tokenize(vocab, s, add_special, parse_special);
    }

    return prompt_tokens;
}

/**
 * break the input "prompt" object into multiple prompt if needed, then tokenize them
 * this supports these cases:
 * - "prompt": "string"
 * - "prompt": [12, 34, 56]
 * - "prompt": [12, 34, "string", 56, 78]
 * and multiple prompts (multi-tasks):
 * - "prompt": ["string1", "string2"]
 * - "prompt": ["string1", [12, 34, 56]]
 * - "prompt": [[12, 34, 56], [78, 90, 12]]
 * - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
 */
static std::vector<llama_tokens> tokenize_input_prompts(const llama_vocab * vocab, const json & json_prompt, bool add_special, bool parse_special) {
    std::vector<llama_tokens> result;
    if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
        // string or mixed
        result.push_back(tokenize_mixed(vocab, json_prompt, add_special, parse_special));
    } else if (json_is_array_of_numbers(json_prompt)) {
        // array of tokens
        result.push_back(json_prompt.get<llama_tokens>());
    } else if (json_prompt.is_array()) {
        // array of prompts
        result.reserve(json_prompt.size());
        for (const auto & p : json_prompt) {
            if (p.is_string() || json_is_array_of_mixed_numbers_strings(p)) {
                result.push_back(tokenize_mixed(vocab, p, add_special, parse_special));
            } else if (json_is_array_of_numbers(p)) {
                // array of tokens
                result.push_back(p.get<llama_tokens>());
            } else {
                throw std::runtime_error("element of \"prompt\" must be a string, an list of tokens, or a list of mixed strings & tokens");
            }
        }
    } else {
        throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts");
    }
    if (result.empty()) {
        throw std::runtime_error("\"prompt\" must not be empty");
    }
    return result;
}

// return the last index of character that can form a valid string
// if the last character is potentially cut in half, return the index before the cut
// if validate_utf8(text) == text.size(), then the whole text is valid utf8
static size_t validate_utf8(const std::string& text) {
    size_t len = text.size();
    if (len == 0) return 0;

    // Check the last few bytes to see if a multi-byte character is cut off
    for (size_t i = 1; i <= 4 && i <= len; ++i) {
        unsigned char c = text[len - i];
        // Check for start of a multi-byte sequence from the end
        if ((c & 0xE0) == 0xC0) {
            // 2-byte character start: 110xxxxx
            // Needs at least 2 bytes
            if (i < 2) return len - i;
        } else if ((c & 0xF0) == 0xE0) {
            // 3-byte character start: 1110xxxx
            // Needs at least 3 bytes
            if (i < 3) return len - i;
        } else if ((c & 0xF8) == 0xF0) {
            // 4-byte character start: 11110xxx
            // Needs at least 4 bytes
            if (i < 4) return len - i;
        }
    }

    // If no cut-off multi-byte character is found, return full length
    return len;
}

//
// template utils
//

// format rerank task: [BOS]query[EOS][SEP]doc[EOS]
static llama_tokens format_rerank(const struct llama_vocab * vocab, const llama_tokens & query, const llama_tokens & doc) {
    llama_tokens result;

    result.reserve(doc.size() + query.size() + 4);
    result.push_back(llama_vocab_bos(vocab));
    result.insert(result.end(), query.begin(), query.end());
    result.push_back(llama_vocab_eos(vocab));
    result.push_back(llama_vocab_sep(vocab));
    result.insert(result.end(), doc.begin(), doc.end());
    result.push_back(llama_vocab_eos(vocab));

    return result;
}

// format infill task
static llama_tokens format_infill(
        const llama_vocab * vocab,
        const json & input_prefix,
        const json & input_suffix,
        const json & input_extra,
        const int n_batch,
        const int n_predict,
        const int n_ctx,
        const bool spm_infill,
        const llama_tokens & tokens_prompt
    ) {
    // TODO: optimize this block by reducing memory allocations and movement

    // use FIM repo-level pattern:
    // ref: https://arxiv.org/pdf/2409.12186
    //
    // [FIM_REP]myproject
    // [FIM_SEP]filename0
    // extra chunk 0
    // [FIM_SEP]filename1
    // extra chunk 1
    // ...
    // [FIM_SEP]filename
    // [FIM_PRE]prefix[FIM_SUF]suffix[FIM_MID]prompt
    //
    llama_tokens extra_tokens;
    extra_tokens.reserve(n_ctx);

    auto tokens_prefix = tokenize_mixed(vocab, input_prefix, false, false);
    auto tokens_suffix = tokenize_mixed(vocab, input_suffix, false, false);

    if (llama_vocab_fim_rep(vocab) != LLAMA_TOKEN_NULL) {
        // TODO: make project name an input
        static const auto k_fim_repo = common_tokenize(vocab, "myproject\n", false, false);

        extra_tokens.push_back(llama_vocab_fim_rep(vocab));
        extra_tokens.insert(extra_tokens.end(), k_fim_repo.begin(), k_fim_repo.end());
    }
    for (const auto & chunk : input_extra) {
        // { "text": string, "filename": string }
        const std::string text     = json_value(chunk, "text",     std::string());
        const std::string filename = json_value(chunk, "filename", std::string("tmp"));

        if (llama_vocab_fim_sep(vocab) != LLAMA_TOKEN_NULL) {
            const auto k_fim_file = common_tokenize(vocab, filename + "\n", false, false);

            extra_tokens.insert(extra_tokens.end(), llama_vocab_fim_sep(vocab));
            extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
        } else {
            // chunk separator in binary form to avoid confusing the AI
            static const char k_chunk_prefix_str[] = {0x0a, 0x0a, 0x2d, 0x2d, 0x2d, 0x20, 0x73, 0x6e, 0x69, 0x70, 0x70, 0x65, 0x74, 0x20, 0x2d, 0x2d, 0x2d, 0x0a, 0x0a, 0x00};
            static const auto k_chunk_prefix_tokens = common_tokenize(vocab, k_chunk_prefix_str, false, false);

            extra_tokens.insert(extra_tokens.end(), k_chunk_prefix_tokens.begin(), k_chunk_prefix_tokens.end());
        }

        const auto chunk_tokens = common_tokenize(vocab, text, false, false);
        extra_tokens.insert(extra_tokens.end(), chunk_tokens.begin(), chunk_tokens.end());
    }

    if (llama_vocab_fim_sep(vocab) != LLAMA_TOKEN_NULL) {
        // TODO: current filename
        static const auto k_fim_file = common_tokenize(vocab, "filename\n", false, false);

        extra_tokens.insert(extra_tokens.end(), llama_vocab_fim_sep(vocab));
        extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
    }

    // for now pick FIM context to fit in a batch (ratio prefix:suffix = 3:1, TODO: configurable?)
    const int n_prefix_take = std::min<int>(tokens_prefix.size(),                3*(n_batch/4));
    const int n_suffix_take = std::min<int>(tokens_suffix.size(), std::max<int>(0, (n_batch/4) - (2 + tokens_prompt.size())));

    SRV_DBG("n_prefix_take = %d, n_suffix_take = %d, total = %d\n", n_prefix_take, n_suffix_take, (n_prefix_take + n_suffix_take));

    // fill the rest of the context with extra chunks
    const int n_extra_take = std::min<int>(std::max<int>(0, n_ctx - (n_batch) - 2*n_predict), extra_tokens.size());

    tokens_prefix.erase(tokens_prefix.begin(), tokens_prefix.begin() + tokens_prefix.size() - n_prefix_take);
    tokens_suffix.resize(n_suffix_take);

    tokens_prefix.insert(tokens_prefix.begin(), llama_vocab_fim_pre(vocab));
    tokens_prefix.insert(tokens_prefix.end(),   tokens_prompt.begin(), tokens_prompt.end());
    tokens_suffix.insert(tokens_suffix.begin(), llama_vocab_fim_suf(vocab));

    auto embd_inp = spm_infill ? tokens_suffix : tokens_prefix;
    auto embd_end = spm_infill ? tokens_prefix : tokens_suffix;

    if (llama_vocab_get_add_bos(vocab)) {
        embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
    }

    SRV_DBG("extra: n_ctx = %d, n_extra_take = %d, n_extra = %d\n", n_ctx, n_extra_take, (int) extra_tokens.size());

    // put the extra context before the FIM prefix
    embd_inp.insert(embd_inp.begin(), extra_tokens.end() - n_extra_take, extra_tokens.end());

    embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
    embd_inp.push_back(llama_vocab_fim_mid(vocab));

    return embd_inp;
}

//
// base64 utils (TODO: move to common in the future)
//

static const std::string base64_chars =
             "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
             "abcdefghijklmnopqrstuvwxyz"
             "0123456789+/";

static inline bool is_base64(uint8_t c) {
    return (isalnum(c) || (c == '+') || (c == '/'));
}

static inline raw_buffer base64_decode(const std::string & encoded_string) {
    int i = 0;
    int j = 0;
    int in_ = 0;

    int in_len = encoded_string.size();

    uint8_t char_array_4[4];
    uint8_t char_array_3[3];

    raw_buffer ret;

    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) {
        char_array_4[i++] = encoded_string[in_]; in_++;
        if (i == 4) {
            for (i = 0; i < 4; i++) {
                char_array_4[i] = base64_chars.find(char_array_4[i]);
            }

            char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
            char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
            char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];

            for (i = 0; (i < 3); i++) {
                ret.push_back(char_array_3[i]);
            }

            i = 0;
        }
    }

    if (i) {
        for (j = i; j < 4; j++) {
            char_array_4[j] = 0;
        }

        for (j = 0; j < 4; j++) {
            char_array_4[j] = base64_chars.find(char_array_4[j]);
        }

        char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
        char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
        char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];

        for (j = 0; j < i - 1; j++) {
            ret.push_back(char_array_3[j]);
        }
    }

    return ret;
}

//
// random string / id
//

static std::string random_string() {
    static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");

    std::random_device rd;
    std::mt19937 generator(rd());

    std::string result(32, ' ');

    for (int i = 0; i < 32; ++i) {
        result[i] = str[generator() % str.size()];
    }

    return result;
}

static std::string gen_chatcmplid() {
    return "chatcmpl-" + random_string();
}

static std::string gen_tool_call_id() {
    return random_string();
}

//
// other common utils
//

static bool ends_with(const std::string & str, const std::string & suffix) {
    return str.size() >= suffix.size() && 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
}

static size_t find_partial_stop_string(const std::string &stop, const std::string &text) {
    if (!text.empty() && !stop.empty()) {
        const char text_last_char = text.back();
        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
            if (stop[char_index] == text_last_char) {
                const std::string current_partial = stop.substr(0, char_index + 1);
                if (ends_with(text, current_partial)) {
                    return text.size() - char_index - 1;
                }
            }
        }
    }

    return std::string::npos;
}

// TODO: reuse llama_detokenize
template <class Iter>
static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
    std::string ret;
    for (; begin != end; ++begin) {
        ret += common_token_to_piece(ctx, *begin);
    }

    return ret;
}

// format incomplete utf-8 multibyte character for output
static std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token) {
    std::string out = token == LLAMA_TOKEN_NULL ? "" : common_token_to_piece(ctx, token);

    // if the size is 1 and first bit is 1, meaning it's a partial character
    //   (size > 1 meaning it's already a known token)
    if (out.size() == 1 && (out[0] & 0x80) == 0x80) {
        std::stringstream ss;
        ss << std::hex << (out[0] & 0xff);
        std::string res(ss.str());
        out = "byte: \\x" + res;
    }

    return out;
}

static bool server_sent_event(httplib::DataSink & sink, const char * event, const json & data) {
    const std::string str =
        std::string(event) + ": " +
        data.dump(-1, ' ', false, json::error_handler_t::replace) +
        "\n\n"; // required by RFC 8895 - A message is terminated by a blank line (two line terminators in a row).

    LOG_DBG("data stream, to_send: %s", str.c_str());

    return sink.write(str.c_str(), str.size());
}

//
// OAI utils
//

static json oaicompat_completion_params_parse(const json & body) {
    json llama_params;

    if (!body.contains("prompt")) {
        throw std::runtime_error("\"prompt\" is required");
    }

    // Handle "stop" field
    if (body.contains("stop") && body.at("stop").is_string()) {
        llama_params["stop"] = json::array({body.at("stop").get<std::string>()});
    } else {
        llama_params["stop"] = json_value(body, "stop", json::array());
    }

    // Handle "n" field
    int n_choices = json_value(body, "n", 1);
    if (n_choices != 1) {
        throw std::runtime_error("Only one completion choice is allowed");
    }

    // Handle "echo" field
    if (json_value(body, "echo", false)) {
        throw std::runtime_error("Only no echo is supported");
    }

    // Params supported by OAI but unsupported by llama.cpp
    static const std::vector<std::string> unsupported_params { "best_of", "suffix" };
    for (const auto & param : unsupported_params) {
        if (body.contains(param)) {
            throw std::runtime_error("Unsupported param: " + param);
        }
    }

    // Copy remaining properties to llama_params
    for (const auto & item : body.items()) {
        // Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens"
        if (!llama_params.contains(item.key()) || item.key() == "n_predict") {
            llama_params[item.key()] = item.value();
        }
    }

    return llama_params;
}

static json oaicompat_completion_params_parse(
    const json & body, /* openai api json semantics */
    bool use_jinja,
    bool prefill_assistant,
    common_reasoning_format reasoning_format,
    const struct common_chat_templates * tmpls,
    bool allow_non_text,
    std::vector<raw_buffer> & out_files)
{
    json llama_params;

    auto tools = json_value(body, "tools", json());
    auto stream = json_value(body, "stream", false);

    if (tools.is_array() && !tools.empty()) {
        if (stream) {
            throw std::runtime_error("Cannot use tools with stream");
        }
        if (!use_jinja) {
            throw std::runtime_error("tools param requires --jinja flag");
        }
    }
    if (!use_jinja) {
        if (body.contains("tool_choice") && !body.at("tool_choice").is_null()) {
            throw std::runtime_error("Unsupported param: tool_choice");
        }
    }

    // Handle "stop" field
    if (body.contains("stop") && body.at("stop").is_string()) {
        llama_params["stop"] = json::array({body.at("stop").get<std::string>()});
    } else {
        llama_params["stop"] = json_value(body, "stop", json::array());
    }

    auto json_schema = json_value(body, "json_schema", json());
    auto grammar = json_value(body, "grammar", std::string());
    if (!json_schema.is_null() && !grammar.empty()) {
        throw std::runtime_error("Cannot use both json_schema and grammar");
    }

    // Handle "response_format" field
    if (body.contains("response_format")) {
        json response_format      = json_value(body, "response_format", json::object());
        std::string response_type = json_value(response_format, "type", std::string());
        if (response_type == "json_object") {
            json_schema = json_value(response_format, "schema", json::object());
        } else if (response_type == "json_schema") {
            auto schema_wrapper = json_value(response_format, "json_schema", json::object());
            json_schema = json_value(schema_wrapper, "schema", json::object());
        } else if (!response_type.empty() && response_type != "text") {
            throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type);
        }
    }

    // get input files
    if (!body.contains("messages")) {
        throw std::runtime_error("'messages' is required");
    }
    json messages = body.at("messages");
    if (!messages.is_array()) {
        throw std::runtime_error("Expected 'messages' to be an array");
    }
    for (auto & msg : messages) {
        std::string role = json_value(msg, "role", std::string());
        if (role != "assistant" && !msg.contains("content")) {
            throw std::runtime_error("All non-assistant messages must contain 'content'");
        }
        if (role == "assistant") {
            if (!msg.contains("content") && !msg.contains("tool_calls")) {
                throw std::runtime_error("Assistant message must contain either 'content' or 'tool_calls'!");
            }
            if (!msg.contains("content")) {
                continue; // avoid errors with no content
            }
        }
        json & content = msg.at("content");
        if (content.is_string() || content.is_null()) {
            continue;
        }

        if (!content.is_array()) {
            throw std::runtime_error("Expected 'content' to be a string or an array");
        }

        for (auto & p : content) {
            std::string type      = json_value(p, "type", std::string());
            json        image_url = json_value(p, "image_url", json::object());
            if (type == "image_url") {
                if (!allow_non_text) {
                    throw std::runtime_error("image input is not supported by this server");
                }

                std::string url = json_value(image_url, "url", std::string());
                if (string_starts_with(url, "http")) {
                    // download remote image
                    // TODO @ngxson : maybe make these params configurable
                    common_remote_params params;
                    params.headers.push_back("User-Agent: llama.cpp/" + build_info);
                    params.max_size = 1024 * 1024 * 10; // 10MB
                    params.timeout  = 10; // seconds
                    SRV_INF("downloading image from '%s'\n", url.c_str());
                    auto res = common_remote_get_content(url, params);
                    if (200 <= res.first && res.first < 300) {
                        SRV_INF("downloaded %ld bytes\n", res.second.size());
                        raw_buffer data;
                        data.insert(data.end(), res.second.begin(), res.second.end());
                        out_files.push_back(data);
                    } else {
                        throw std::runtime_error("Failed to download image");
                    }

                } else {
                    // try to decode base64 image
                    std::vector<std::string> parts = string_split<std::string>(url, /*separator*/ ',');
                    if (parts.size() != 2) {
                        throw std::runtime_error("Invalid image_url.url value");
                    } else if (!string_starts_with(parts[0], "data:image/")) {
                        throw std::runtime_error("Invalid image_url.url format: " + parts[0]);
                    } else if (!string_ends_with(parts[0], "base64")) {
                        throw std::runtime_error("image_url.url must be base64 encoded");
                    } else {
                        auto base64_data = parts[1];
                        auto decoded_data = base64_decode(base64_data);
                        out_files.push_back(decoded_data);
                    }
                }

                // replace this chunk with a marker
                p["type"] = "text";
                p["text"] = MTMD_DEFAULT_IMAGE_MARKER;
                p.erase("image_url");
            }
        }
    }

    common_chat_templates_inputs inputs;
    inputs.messages              = common_chat_msgs_parse_oaicompat(messages);
    inputs.tools                 = common_chat_tools_parse_oaicompat(tools);
    inputs.tool_choice           = common_chat_tool_choice_parse_oaicompat(json_value(body, "tool_choice", std::string("auto")));
    inputs.json_schema           = json_schema.is_null() ? "" : json_schema.dump();
    inputs.grammar               = grammar;
    inputs.add_generation_prompt = json_value(body, "add_generation_prompt", true);
    inputs.use_jinja             = use_jinja;
    inputs.parallel_tool_calls   = json_value(body, "parallel_tool_calls", false);
    inputs.extract_reasoning     = reasoning_format != COMMON_REASONING_FORMAT_NONE;
    inputs.add_generation_prompt = json_value(body, "add_generation_prompt", true);
    if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && body.contains("grammar")) {
        throw std::runtime_error("Cannot use custom grammar constraints with tools.");
    }

    // if the assistant message appears at the end of list, we do not add end-of-turn token
    // for ex. this can be useful to modify the reasoning process in reasoning models
    bool prefill_assistant_message = !inputs.messages.empty() && inputs.messages.back().role == "assistant" && prefill_assistant;
    common_chat_msg last_message;
    if (prefill_assistant_message) {
        last_message = inputs.messages.back();
        inputs.messages.pop_back();

        /* sanity check, max one assistant message at the end of the list */
        if (!inputs.messages.empty() && inputs.messages.back().role == "assistant"){
            throw std::runtime_error("Cannot have 2 or more assistant messages at the end of the list.");
        }

        inputs.extract_reasoning = false;
        inputs.add_generation_prompt = true;
    }

    // Apply chat template to the list of messages
    auto chat_params = common_chat_templates_apply(tmpls, inputs);

    /* Append assistant prefilled message */
    if (prefill_assistant_message) {
         chat_params.prompt += last_message.content;
    }

    llama_params["chat_format"]      = static_cast<int>(chat_params.format);
    llama_params["prompt"]           = chat_params.prompt;
    if (!chat_params.grammar.empty()) {
        llama_params["grammar"] = chat_params.grammar;
    }
    llama_params["grammar_lazy"]     = chat_params.grammar_lazy;
    auto grammar_triggers = json::array();
    for (const auto & trigger : chat_params.grammar_triggers) {
        server_grammar_trigger ct(trigger);
        grammar_triggers.push_back(ct.to_json());
    }
    llama_params["grammar_triggers"] = grammar_triggers;
    llama_params["preserved_tokens"] = chat_params.preserved_tokens;
    for (const auto & stop : chat_params.additional_stops) {
        llama_params["stop"].push_back(stop);
    }

    // Handle "n" field
    int n_choices = json_value(body, "n", 1);
    if (n_choices != 1) {
        throw std::runtime_error("Only one completion choice is allowed");
    }

    // Handle "logprobs" field
    // TODO: The response format of this option is not yet OAI-compatible, but seems like no one really using it; We may need to fix it in the future
    if (json_value(body, "logprobs", false)) {
        llama_params["n_probs"] = json_value(body, "top_logprobs", 20);
    } else if (body.contains("top_logprobs") && !body.at("top_logprobs").is_null()) {
        throw std::runtime_error("top_logprobs requires logprobs to be set to true");
    }

    // Copy remaining properties to llama_params
    // This allows user to use llama.cpp-specific params like "mirostat", ... via OAI endpoint.
    // See "launch_slot_with_task()" for a complete list of params supported by llama.cpp
    for (const auto & item : body.items()) {
        // Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens"
        if (!llama_params.contains(item.key()) || item.key() == "n_predict") {
            llama_params[item.key()] = item.value();
        }
    }

    return llama_params;
}

static json format_embeddings_response_oaicompat(const json & request, const json & embeddings, bool use_base64 = false) {
    json data = json::array();
    int32_t n_tokens = 0;
    int i = 0;
    for (const auto & elem : embeddings) {
        json embedding_obj;

        if (use_base64) {
            const auto& vec = json_value(elem, "embedding", json::array()).get<std::vector<float>>();
            const char* data_ptr = reinterpret_cast<const char*>(vec.data());
            size_t data_size = vec.size() * sizeof(float);
            embedding_obj = {
                {"embedding", base64::encode(data_ptr, data_size)},
                {"index", i++},
                {"object", "embedding"},
                {"encoding_format", "base64"}
            };
        } else {
            embedding_obj = {
                {"embedding", json_value(elem, "embedding", json::array())},
                {"index", i++},
                {"object", "embedding"}
            };
        }
        data.push_back(embedding_obj);

        n_tokens += json_value(elem, "tokens_evaluated", 0);
    }

    json res = json {
        {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
        {"object", "list"},
        {"usage", json {
            {"prompt_tokens", n_tokens},
            {"total_tokens", n_tokens}
        }},
        {"data", data}
    };

    return res;
}

static json format_response_rerank(
        const json & request,
        const json & ranks,
        bool is_tei_format,
        std::vector<std::string> & texts) {
    json res;
    if (is_tei_format) {
        // TEI response format
        res = json::array();
        bool return_text = json_value(request, "return_text", false);
        for (const auto & rank : ranks) {
            int index = json_value(rank, "index", 0);
            json elem = json{
                {"index", index},
                {"score", json_value(rank, "score", 0.0)},
            };
            if (return_text) {
                elem["text"] = std::move(texts[index]);
            }
            res.push_back(elem);
        }
    } else {
        // Jina response format
        json results = json::array();
        int32_t n_tokens = 0;
        for (const auto & rank : ranks) {
            results.push_back(json{
                {"index",           json_value(rank, "index", 0)},
                {"relevance_score", json_value(rank, "score", 0.0)},
            });

            n_tokens += json_value(rank, "tokens_evaluated", 0);
        }

        res = json{
            {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
            {"object", "list"},
            {"usage", json{
                {"prompt_tokens", n_tokens},
                {"total_tokens", n_tokens}
            }},
            {"results", results}
        };
    }

    return res;
}

static bool is_valid_utf8(const std::string & str) {
    const unsigned char* bytes = reinterpret_cast<const unsigned char*>(str.data());
    const unsigned char* end = bytes + str.length();

    while (bytes < end) {
        if (*bytes <= 0x7F) {
            // 1-byte sequence (0xxxxxxx)
            bytes++;
        } else if ((*bytes & 0xE0) == 0xC0) {
            // 2-byte sequence (110xxxxx 10xxxxxx)
            if (end - bytes < 2 || (bytes[1] & 0xC0) != 0x80)
                return false;
            bytes += 2;
        } else if ((*bytes & 0xF0) == 0xE0) {
            // 3-byte sequence (1110xxxx 10xxxxxx 10xxxxxx)
            if (end - bytes < 3 || (bytes[1] & 0xC0) != 0x80 || (bytes[2] & 0xC0) != 0x80)
                return false;
            bytes += 3;
        } else if ((*bytes & 0xF8) == 0xF0) {
            // 4-byte sequence (11110xxx 10xxxxxx 10xxxxxx 10xxxxxx)
            if (end - bytes < 4 || (bytes[1] & 0xC0) != 0x80 ||
                (bytes[2] & 0xC0) != 0x80 || (bytes[3] & 0xC0) != 0x80)
                return false;
            bytes += 4;
        } else {
            // Invalid UTF-8 lead byte
            return false;
        }
    }

    return true;
}

static json format_tokenizer_response(const json & tokens) {
    return json {
        {"tokens", tokens}
    };
}

static json format_detokenized_response(const std::string & content) {
    return json {
        {"content", content}
    };
}

static json format_logit_bias(const std::vector<llama_logit_bias> & logit_bias) {
    json data = json::array();
    for (const auto & lb : logit_bias) {
        data.push_back(json{
            {"bias", lb.bias},
            {"token", lb.token},
        });
    }
    return data;
}

static std::string safe_json_to_str(const json & data) {
    return data.dump(-1, ' ', false, json::error_handler_t::replace);
}

static std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx) {
    std::vector<llama_token_data> cur;
    const auto * logits = llama_get_logits_ith(ctx, idx);

    const llama_model * model = llama_get_model(ctx);
    const llama_vocab * vocab = llama_model_get_vocab(model);

    const int n_vocab = llama_vocab_n_tokens(vocab);

    cur.resize(n_vocab);
    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
        cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
    }

    // sort tokens by logits
    std::sort(cur.begin(), cur.end(), [](const llama_token_data & a, const llama_token_data & b) {
        return a.logit > b.logit;
    });

    // apply softmax
    float max_l = cur[0].logit;
    float cum_sum = 0.0f;
    for (size_t i = 0; i < cur.size(); ++i) {
        float p = expf(cur[i].logit - max_l);
        cur[i].p = p;
        cum_sum += p;
    }
    for (size_t i = 0; i < cur.size(); ++i) {
        cur[i].p /= cum_sum;
    }

    return cur;
}

static bool are_lora_equal(
        const std::vector<common_adapter_lora_info> & l1,
        const std::vector<common_adapter_lora_info> & l2) {
    if (l1.size() != l2.size()) {
        return false;
    }
    for (size_t i = 0; i < l1.size(); ++i) {
        // we don't check lora.path to reduce the time complexity
        if (l1[i].scale != l2[i].scale || l1[i].ptr != l2[i].ptr) {
            return false;
        }
    }
    return true;
}

// parse lora config from JSON request, returned a copy of lora_base with updated scale
static std::vector<common_adapter_lora_info> parse_lora_request(
        const std::vector<common_adapter_lora_info> & lora_base,
        const json & data) {
    std::vector<common_adapter_lora_info> lora(lora_base);
    int max_idx = lora.size();

    // clear existing value
    for (auto & entry : lora) {
        entry.scale = 0.0f;
    }

    // set value
    for (const auto & entry : data) {
        int id      = json_value(entry, "id", -1);
        float scale = json_value(entry, "scale", 0.0f);
        if (0 <= id && id < max_idx) {
            lora[id].scale = scale;
        } else {
            throw std::runtime_error("invalid adapter id");
        }
    }

    return lora;
}

//
// utils for interacting with libmtmd
// (may need to refactor in near future)
//

/**
 * server_tokens is a helper to manage the input tokens and image for the server.
 * it is made this way to simplify the logic of KV cache management.
 */
struct server_tokens {
    bool has_mtmd = false;

private: // disallow accessing these members directly, risking out-of-sync

    // map a **start** position in tokens to the image chunk
    std::unordered_map<llama_pos, mtmd::input_chunk_ptr> map_pos_to_image;

    // list of tokens
    // it can include LLAMA_TOKEN_NULL, which is used to indicate a token that is not a text token
    // a mtmd_input_chunk can occupy multiple tokens, one llama_token per **position**
    // important: for models using mrope, an image can contain multiple tokens but will use only one **position**
    llama_tokens tokens;

    // for ex. with input of 5 text tokens and 2 images:
    //      [0] [1] [2] [3] [4] [img0] [img0] [img0] [img1] [img1]
    // pos  0   1   2   3   4   5      6      7      8      9
    // map_pos_to_image will contain: {5, img0}, {8, img1}

public:
    server_tokens() = default;
    ~server_tokens() = default;

    // Prevent copying
    server_tokens(const server_tokens&) = delete;
    server_tokens& operator=(const server_tokens&) = delete;

    // Allow moving (usually implicitly generated if members are movable)
    server_tokens(server_tokens&&) = default;
    server_tokens& operator=(server_tokens&&) = default;

    // Allow accessing elements using [] operator
    llama_token operator[](size_t index) { return tokens[index]; }
    const llama_token& operator[](size_t index) const { return tokens[index]; }

    server_tokens(mtmd::input_chunks & mtmd_chunks, bool has_mtmd) : has_mtmd(has_mtmd) {
        for (size_t i = 0; i < mtmd_chunks.size(); ++i) {
            push_back(mtmd_chunks[i]);
        }
    }

    server_tokens(llama_tokens & tokens, bool has_mtmd) : has_mtmd(has_mtmd), tokens(tokens) {}

    // for debugging
    std::string str() const {
        std::ostringstream oss;
        oss << "tokens: ";
        for (const auto & t : tokens) {
            if (t == LLAMA_TOKEN_NULL) {
                oss << "<embd> ";
            } else {
                oss << t << " ";
            }
        }
        oss << "\n";
        oss << "image pos: ";
        for (const auto & it : map_pos_to_image) {
            oss << it.first << ", ";
        }
        return oss.str();
    }

    const mtmd::input_chunk_ptr & find_chunk(llama_pos pos) const {
        auto it = map_pos_to_image.find(pos);
        if (it != map_pos_to_image.end()) {
            return it->second;
        } else {
            throw std::runtime_error("Chunk not found");
        }
    }

    void push_back(llama_token tok) {
        if (tok == LLAMA_TOKEN_NULL) {
            throw std::runtime_error("Invalid token");
        }
        tokens.emplace_back(tok);
    }

    // will create a copy of the chunk if it contains non-text data
    void push_back(const mtmd_input_chunk * chunk) {
        auto type = mtmd_input_chunk_get_type(chunk);
        if (type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
            GGML_ASSERT(has_mtmd);
            auto img_tokens = mtmd_input_chunk_get_tokens_image(chunk);
            const int n_pos = mtmd_image_tokens_get_n_pos(img_tokens);
            llama_pos start_pos = tokens.size();
            for (int i = 0; i < n_pos; ++i) {
                tokens.emplace_back(LLAMA_TOKEN_NULL);
            }
            mtmd::input_chunk_ptr new_chunk(mtmd_input_chunk_copy(chunk));
            map_pos_to_image[start_pos] = std::move(new_chunk);
        } else if (type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
            size_t n_tokens;
            auto text_tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens);
            for (size_t i = 0; i < n_tokens; ++i) {
                push_back(text_tokens[i]);
            }
        } else {
            GGML_ABORT("Invalid chunk type");
        }
    }

    // for compatibility with context shift and prompt truncation
    void insert(const llama_tokens & inp_tokens) {
        GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled
        tokens.insert(tokens.end(), inp_tokens.begin(), inp_tokens.end());
    }

    // for compatibility with speculative decoding, ctx shift, slot save/load
    const llama_tokens & get_text_tokens() const {
        GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled
        return tokens;
    }

    // for compatibility with speculative decoding
    void set_token(llama_pos pos, llama_token id) {
        GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled
        tokens[pos] = id;
    }

    size_t size() const {
        return tokens.size();
    }

    bool empty() const {
        return tokens.empty();
    }

    void clear() {
        tokens.clear();
    }

    void keep_first(size_t n) {
        GGML_ASSERT(n <= tokens.size());
        if (has_mtmd) {
            // we throw an error if we try to remove a token in the middle of an image
            // for ex. with input of 5 text tokens and 2 images:
            //    [0] [1] [2] [3] [4] [img0] [img0] [img0] [img1] [img1]
            // n  1   2   3   4   5   6      7      8      9      10
            // allowed to resize      ^                    ^
            // disallowed to resize          ^      ^             ^
            if (n > 0) {
                llama_token last_token = tokens[n - 1];
                // make sure we never remove tokens in the middle of an image
                if (last_token == LLAMA_TOKEN_NULL) {
                    find_chunk(n - 1); // will throw an error if the token is not begin-of-chunk
                }
            }
            // remove all image chunks that are not used anymore
            for (auto it = map_pos_to_image.begin(); it != map_pos_to_image.end(); ) {
                llama_pos pos = it->first;
                if (pos >= (llama_pos)n) {
                    it = map_pos_to_image.erase(it);
                } else {
                    ++it;
                }
            }
        }
        tokens.resize(n);
    }

    std::string detokenize(const llama_context * ctx, bool special) const {
        llama_tokens text_tokens;
        text_tokens.reserve(tokens.size());
        for (const auto & t : tokens) {
            if (t != LLAMA_TOKEN_NULL) {
                text_tokens.push_back(t);
            }
        }
        return common_detokenize(ctx, text_tokens, special);
    }

    size_t get_common_prefix(const server_tokens & b) const {
        size_t max_idx = std::min(tokens.size(), b.tokens.size());
        for (size_t i = 0; i < max_idx; ++i) {
            auto & ai =   tokens[i];
            auto & bi = b.tokens[i];

            if (ai == LLAMA_TOKEN_NULL && bi == LLAMA_TOKEN_NULL) {
                GGML_ASSERT(has_mtmd);
                const auto & a_chunk =   find_chunk(i);
                const auto & b_chunk = b.find_chunk(i);
                GGML_ASSERT(a_chunk && b_chunk);
                const auto * a_img = mtmd_input_chunk_get_tokens_image(a_chunk.get());
                const auto * b_img = mtmd_input_chunk_get_tokens_image(b_chunk.get());
                std::string ai_id  = mtmd_image_tokens_get_id(a_img);
                std::string bi_id  = mtmd_image_tokens_get_id(b_img);
                size_t a_pos       = mtmd_image_tokens_get_n_pos(a_img);
                size_t b_pos       = mtmd_image_tokens_get_n_pos(b_img);
                if (ai_id == bi_id && a_pos == b_pos) {
                    GGML_ASSERT(a_pos > 0 && "Invalid image token"); // should never happen
                    i += a_pos - 1; // will be +1 by the for loop
                    continue;
                } else {
                    return i;
                }
            } else if (ai == bi) {
                continue;
            } else {
                return i;
            }
        }
        return max_idx; // all tokens are equal
    }

    // make sure all text tokens are within the vocab range
    bool validate(const struct llama_context * ctx) const {
        const llama_model * model = llama_get_model(ctx);
        const llama_vocab * vocab = llama_model_get_vocab(model);
        const int32_t n_vocab = llama_vocab_n_tokens(vocab);

        for (size_t i = 0; i < tokens.size(); ++i) {
            auto & t = tokens[i];
            if (t == LLAMA_TOKEN_NULL) {
                try {
                    const auto & chunk = find_chunk(i);
                    const auto * img_tokens = mtmd_input_chunk_get_tokens_image(chunk.get());
                    size_t n_pos = mtmd_image_tokens_get_n_pos(img_tokens);
                    i += n_pos - 1; // will be +1 by the for loop
                } catch (const std::exception & e) {
                    return false;
                }
            } else if (t < 0 || t >= n_vocab) {
                return false;
            }
        }
        return true;
    }

    // encode and decode the image chunk
    int32_t process_chunk(
                llama_context * ctx,
                mtmd_context * mctx,
                llama_pos n_past,
                int32_t seq_id,
                llama_pos & n_pos_out) {
        auto it = map_pos_to_image.find(n_past);
        if (it == map_pos_to_image.end()) {
            throw std::runtime_error("Chunk not found");
        }
        SRV_INF("%s\n", "processing image...");
        int32_t n_batch = llama_n_batch(ctx);
        int64_t t0 = ggml_time_ms();
        llama_pos new_n_past = n_past;
        int32_t result = mtmd_helper_eval_chunk_single(mctx, ctx,
            it->second.get(), // chunk
            n_past,
            seq_id,
            n_batch,
            true, // logits last
            &new_n_past);
        SRV_INF("image processed in %" PRId64 " ms\n", ggml_time_ms() - t0);
        if (result != 0) {
            LOG_ERR("mtmd_helper_eval failed with status %d", result);
            n_pos_out = n_past;
            return result;
        }
        n_pos_out = new_n_past;
        return 0;
    }
};

// Computes FNV-1a hash of the data
static std::string fnv_hash(const uint8_t * data, size_t len) {
    const uint64_t fnv_prime = 0x100000001b3ULL;
    uint64_t hash = 0xcbf29ce484222325ULL;

    for (size_t i = 0; i < len; ++i) {
        hash ^= data[i];
        hash *= fnv_prime;
    }
    return std::to_string(hash);
}