char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Aldehir Rojascommon : implement common_schema internal representation for JSON schemas (#28736)acecd5603

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
3.1 KiB76 linesraw
1#include "parsers.h"
2
3common_chat_params common_chat_params_init_gigachat_v3(
4        const common_chat_template & tmpl,
5        const autoparser::generation_params & inputs) {
6
7    common_chat_params data;
8
9    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
10    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
11    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
12    data.supports_thinking = false;
13    data.preserved_tokens  = {
14        "<|message_sep|>\n\n",
15        "<|role_sep|>\n",
16    };
17
18    if (inputs.has_continuation()) {
19        const auto & msg = inputs.continue_msg;
20        data.generation_prompt = "assistant<|role_sep|>\n" + msg.render_content();
21        data.prompt += data.generation_prompt;
22    }
23
24    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
25    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
26    const auto *tool_call_start_prefix = "<|message_sep|>\n\nfunction call<|role_sep|>\n";
27
28    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
29        auto ret = p.eps();
30        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
31            // Build a choice of all available tools
32            auto tool_choice = p.choice();
33            for (const auto & tool : inputs.tools) {
34                const auto & function = tool.at("function");
35                std::string name = function.at("name");
36                const auto  schema = common_chat_tool_parameters(function);
37
38                auto tool_name = p.json_member("name", "\"" + p.tool_name(p.literal(name)) + "\"");
39                auto tool_args = p.json_member("arguments", p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)));
40
41                auto tool_open = p.tool_open(p.literal("{") << tool_name);
42
43                tool_choice |= p.rule("tool-" + name, tool_open << "," << tool_args << "}");
44            }
45
46            // Define the tool call structure
47            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
48            auto max_calls = 1; // parallel toolcalls are not supported
49            auto tool_call = p.rule("tool-call", p.literal(tool_call_start_prefix) + tool_choice);
50            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(tool_call, /* min = */ min_calls, /* max = */ max_calls));
51
52            ret = p.content(p.until("<|message_sep|>\n\n")) << tool_calls;
53        } else {
54            // Content only parser
55            include_grammar = false;
56            ret = p.content(p.rest());
57        }
58
59        return p.literal("assistant<|role_sep|>\n") + ret;
60    });
61
62    data.parser = parser.save();
63
64    if (include_grammar) {
65        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
66
67        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
68            parser.build_grammar(builder, data.grammar_lazy);
69        });
70
71        data.grammar_triggers = {
72            {COMMON_GRAMMAR_TRIGGER_TYPE_WORD, tool_call_start_prefix}
73        };
74    }
75    return data;
76}