Function llama_grammar_apply_impl

smallthinker/src/llama-grammar.cpp:1118–1159

}

void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_data_array * cur_p) {
    GGML_ASSERT(grammar.vocab != nullptr);

    if (grammar.awaiting_trigger) {
        return;
    }

    bool allow_eog = false;
    for (const auto & stack : grammar.stacks) {
        if (stack.empty()) {
            allow_eog = true;
            break;
        }
    }

    std::vector<std::pair<std::vector<uint32_t>, llama_partial_utf8>> candidates_decoded;
    candidates_decoded.reserve(cur_p->size);

    llama_grammar_candidates candidates_grammar;
    candidates_grammar.reserve(cur_p->size);

    for (size_t i = 0; i < cur_p->size; ++i) {
        const llama_token id      = cur_p->data[i].id;
        const std::string & piece = grammar.vocab->token_to_piece(id);

        if (grammar.vocab->is_eog(id)) {
            if (!allow_eog) {
                cur_p->data[i].logit = -INFINITY;
            }
        } else if (piece.empty() || piece[0] == 0) {
            cur_p->data[i].logit = -INFINITY;
        } else {
            candidates_decoded.push_back(decode_utf8(piece, grammar.partial_utf8));
            candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second });
        }
    }

    const auto rejects = llama_grammar_reject_candidates(grammar.rules, grammar.stacks, candidates_grammar);
    for (const auto & reject : rejects) {
        cur_p->data[reject.index].logit = -INFINITY;
    }
}

void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token) {
    GGML_ASSERT(grammar.vocab != nullptr);