microtorch::chat::detail Namespace Reference

Functions
std::string	dump_py (const json &x)
std::string	jstr (const std::string &s)
std::string	go (const std::string &s)
void	paperkiln_only (const Template &t, const char *what)
std::string	jinja_tools (const Template &t)
std::string	modelfile_tools (const std::string &gguf_name, const Template &t)
std::string	sha256_hex (const std::string &msg)

Function Documentation

dump_py()

std::string microtorch::chat::detail::dump_py ( const json &  x )

inline

Definition at line 547 of file chat_template.h.

                                        {
    if (x.is_object()) {
        std::string s = "{";
        bool first = true;
        for (auto it = x.begin(); it != x.end(); ++it) {
            if (!first) s += ", ";
            first = false;
            s += json(it.key()).dump() + ": " + dump_py(it.value());
        }
        return s + "}";
    }
    if (x.is_array()) {
        std::string s = "[";
        for (size_t i = 0; i < x.size(); ++i) s += (i ? ", " : "") + dump_py(x[i]);
        return s + "]";
    }
    if (x.is_number_float())
        throw Error("a tool definition holds a non-integer number (refused: its text is not portable)");
    return x.dump();
}

References dump_py(), and j.

Referenced by dump_py(), and microtorch::chat::tool_json().

go()

std::string microtorch::chat::detail::go ( const std::string &  s )

inline

Definition at line 743 of file chat_template.h.

                                        {
    std::string o;
    for (size_t i = 0; i < s.size(); ++i) {
        if (s.compare(i, 2, "{{") == 0) {
            o += "{{ \"{{\" }}";
            ++i;
        } else {
            o += s[i];
        }
    }
    return o;
}

References go(), and j.

Referenced by go(), and modelfile_tools().

jinja_tools()

std::string microtorch::chat::detail::jinja_tools ( const Template &  t )

inline

Definition at line 764 of file chat_template.h.

                                                {
    std::string j;
    if (!t.bos.empty()) j += "{{- " + jstr(t.bos) + " -}}";
    j += "{%- if messages and messages[0]['role'] == 'system' -%}"
         "{%- set sys = messages[0]['content'] -%}{%- set msgs = messages[1:] -%}"
         "{%- else -%}{%- set sys = none -%}{%- set msgs = messages -%}{%- endif -%}";
    if (t.has_tools)
        j += "{%- if tools -%}{%- set ns = namespace(d='') -%}{%- for x in tools -%}"
             "{%- if not loop.first -%}{%- set ns.d = ns.d + " + jstr(t.tools_sep) + " -%}{%- endif -%}"
             "{%- set ns.d = ns.d + (x | tojson) -%}{%- endfor -%}"
             "{%- if sys -%}{%- set sys = sys + " + jstr(t.tools_join) + " + ns.d -%}"
             "{%- else -%}{%- set sys = ns.d -%}{%- endif -%}{%- endif -%}";
    j += "{%- if sys is not none -%}{{- " + jstr(t.system_open) + " + sys + " + jstr(t.system_close) +
         " -}}{%- endif -%}";
    j += "{%- for m in msgs -%}"
         "{%- if m['role'] == 'user' -%}{{- " + jstr(t.user_open) + " + m['content'] + " + jstr(t.user_close) + " -}}"
         "{%- elif m['role'] == 'assistant' -%}{{- " + jstr(t.assistant_open) + " + m['content'] -}}"
         "{%- for c in (m['tool_calls'] or []) -%}{{- " + jstr(t.tool_call_open) + " + c['function']['name'] + " +
         jstr(t.tool_call_join) + " -}}"
         "{%- if c['function']['arguments'] is string -%}{{- c['function']['arguments'] -}}"
         "{%- else -%}{{- c['function']['arguments'] | tojson -}}{%- endif -%}{%- endfor -%}"
         "{{- " + jstr(t.assistant_close) + " -}}"
         "{%- elif m['role'] == 'tool' -%}{{- " + jstr(t.tool_open) + " + m['content'] + " + jstr(t.tool_close) + " -}}"
         "{%- else -%}{{ raise_exception('roles: one leading system message, then user, assistant and tool') }}"
         "{%- endif -%}{%- endfor -%}"
         "{%- if add_generation_prompt -%}{{- " + jstr(t.generation_prompt) + " -}}{%- endif -%}";
    return j;
}

References j, jinja_tools(), and jstr().

Referenced by jinja_tools().

jstr()

std::string microtorch::chat::detail::jstr ( const std::string &  s )

inline

Definition at line 733 of file chat_template.h.

                                          {
    std::string o = "'";
    for (char c : s) {
        if (c == '\\') o += "\\\\";
        else if (c == '\'') o += "\\'";
        else if (c == '\n') o += "\\n";
        else o += c;
    }
    return o + "'";
}

References c, j, and jstr().

Referenced by jinja_tools(), and jstr().

modelfile_tools()

std::string microtorch::chat::detail::modelfile_tools ( const std::string &  gguf_name, const Template &  t  )

inline

Definition at line 794 of file chat_template.h.

                                                                                {
    std::string head = go(t.bos);
    head += "{{ if or .System .Tools }}" + go(t.system_open) + "{{ if .System }}{{ .System }}{{ end }}"
            "{{ if .Tools }}{{ if .System }}" + go(t.tools_join) + "{{ end }}"
            "{{ range $i, $t := .Tools }}{{ if $i }}" + go(t.tools_sep) + "{{ end }}{{ json $t }}{{ end }}"
            "{{ end }}" + go(t.system_close) + "{{ end }}";
    const std::string msgs =
        "{{ range .Messages }}{{ if eq .Role \"user\" }}" + go(t.user_open) + "{{ .Content }}" + go(t.user_close) +
        "{{ else if eq .Role \"assistant\" }}" + go(t.assistant_open) + "{{ .Content }}"
        "{{ range .ToolCalls }}" + go(t.tool_call_open) + "{{ .Function.Name }}" + go(t.tool_call_join) +
        "{{ json .Function.Arguments }}{{ end }}" + go(t.assistant_close) +
        "{{ else if eq .Role \"tool\" }}" + go(t.tool_open) + "{{ .Content }}" + go(t.tool_close) +
        "{{ end }}{{ end }}" + go(t.generation_prompt);
    std::string s = "# Ollama Modelfile for " + gguf_name + ": chat template " + t.id() +
                    " (LOCAL_LLM_STACK.md section 3, point 3).\nFROM ./" + gguf_name + "\nTEMPLATE \"\"\"" + head +
                    msgs + "\"\"\"\n";
    std::vector<std::string> stops = t.stop_eos;
    stops.insert(stops.end(), t.stop_turn.begin(), t.stop_turn.end());
    for (const auto& x : stops) s += "PARAMETER stop " + json(x).dump() + "\n";
    return s;
}

References go(), j, and modelfile_tools().

Referenced by modelfile_tools().

paperkiln_only()

void microtorch::chat::detail::paperkiln_only ( const Template &  t, const char *  what  )

inline

Definition at line 755 of file chat_template.h.

                                                                {
    if (t.name != "paperkiln") throw Error(std::string(what) + " writes paperkiln records only");
}

References j, and paperkiln_only().

Referenced by paperkiln_only().

sha256_hex()

std::string microtorch::chat::detail::sha256_hex ( const std::string &  msg )

inline

Definition at line 871 of file chat_template.h.

                                                  {
    static const uint32_t k[64] = {
        0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5, 0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
        0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3, 0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
        0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc, 0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
        0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7, 0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
        0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13, 0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
        0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3, 0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
        0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5, 0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
        0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208, 0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2};
    uint32_t h[8] = {0x6a09e667, 0xbb67ae85, 0x3c6ef372, 0xa54ff53a, 0x510e527f, 0x9b05688c, 0x1f83d9ab, 0x5be0cd19};
    std::string m = msg;
    const uint64_t bits = static_cast<uint64_t>(msg.size()) * 8;
    m += static_cast<char>(0x80);
    while (m.size() % 64 != 56) m += '\0';
    for (int i = 7; i >= 0; --i) m += static_cast<char>((bits >> (8 * i)) & 0xff);
    auto rotr = [](uint32_t x, int n) { return (x >> n) | (x << (32 - n)); };
    for (size_t off = 0; off < m.size(); off += 64) {
        uint32_t w[64];
        for (int i = 0; i < 16; ++i)
            w[i] = (uint32_t(uint8_t(m[off + 4 * i])) << 24) | (uint32_t(uint8_t(m[off + 4 * i + 1])) << 16) |
                   (uint32_t(uint8_t(m[off + 4 * i + 2])) << 8) | uint32_t(uint8_t(m[off + 4 * i + 3]));
        for (int i = 16; i < 64; ++i) {
            const uint32_t s0 = rotr(w[i - 15], 7) ^ rotr(w[i - 15], 18) ^ (w[i - 15] >> 3);
            const uint32_t s1 = rotr(w[i - 2], 17) ^ rotr(w[i - 2], 19) ^ (w[i - 2] >> 10);
            w[i] = w[i - 16] + s0 + w[i - 7] + s1;
        }
        uint32_t a = h[0], b = h[1], c = h[2], d = h[3], e = h[4], f = h[5], g = h[6], hh = h[7];
        for (int i = 0; i < 64; ++i) {
            const uint32_t t1 = hh + (rotr(e, 6) ^ rotr(e, 11) ^ rotr(e, 25)) + ((e & f) ^ (~e & g)) + k[i] + w[i];
            const uint32_t t2 = (rotr(a, 2) ^ rotr(a, 13) ^ rotr(a, 22)) + ((a & b) ^ (a & c) ^ (b & c));
            hh = g;
            g = f;
            f = e;
            e = d + t1;
            d = c;
            c = b;
            b = a;
            a = t1 + t2;
        }
        h[0] += a; h[1] += b; h[2] += c; h[3] += d; h[4] += e; h[5] += f; h[6] += g; h[7] += hh;
    }
    char buf[65];
    for (int i = 0; i < 8; ++i) std::snprintf(buf + 8 * i, 9, "%08x", h[i]);
    return std::string(buf, 64);
}

References a, c, j, k, n, off, and sha256_hex().

Referenced by sha256_hex().