Method split_seq

subprojects/llama.cpp/src/llama-batch.cpp:611–651

}

llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) {
    // find the first unused token
    uint32_t cur_idx = 0;
    while (cur_idx < used.size() && used[cur_idx]) {
        ++cur_idx;
    }

    // we are done
    if (cur_idx >= used.size()) {
        return {};
    }

    // this is the starting sequence set
    // we allow adding tokens only if their sequence set is a subset of the current sequence set
    auto cur_seq_set = seq_set[cur_idx];

    std::vector<int32_t> idxs;

    while (true) {
        idxs.push_back(cur_idx);

        used[cur_idx] = true;
        ++n_used;

        if (idxs.size() >= n_ubatch) {
            break;
        }

        do {
            ++cur_idx;
        } while (cur_idx < get_n_tokens() && (used[cur_idx] || ((cur_seq_set & seq_set[cur_idx]) != seq_set[cur_idx])));

        if (cur_idx == get_n_tokens()) {
            break;
        }

        cur_seq_set = seq_set[cur_idx];
    }

    return ubatch_add(idxs, 1, true);
}

void llama_batch_allocr::clear() {
    n_outputs = 0;