Method state_read_data

smallthinker/src/llama-context.cpp:1832–1919

}

size_t llama_context::state_read_data(llama_io_read_i & io) {
    LLAMA_LOG_DEBUG("%s: reading state\n", __func__);

    // read model info
    {
        LLAMA_LOG_DEBUG("%s: - reading model info\n", __func__);

        const std::string cur_arch_str = llm_arch_name(model.arch);

        std::string arch_str;
        io.read_string(arch_str);
        if (cur_arch_str != arch_str) {
            throw std::runtime_error(format("wrong model arch: '%s' instead of '%s'", arch_str.c_str(), cur_arch_str.c_str()));
        }
        // TODO: add more info which needs to be identical but which is not verified otherwise
    }

    // read output ids
    {
        LLAMA_LOG_DEBUG("%s: - reading output ids\n", __func__);

        auto n_outputs = this->n_outputs;
        io.read_to(&n_outputs, sizeof(n_outputs));

        if (n_outputs > output_reserve(n_outputs)) {
            throw std::runtime_error("could not reserve outputs");
        }

        std::vector<int32_t> output_pos;

        if (n_outputs) {
            output_pos.resize(n_outputs);
            io.read_to(output_pos.data(), n_outputs * sizeof(int32_t));

            for (int32_t i = 0; i < (int32_t) output_pos.size(); ++i) {
                int32_t id = output_pos[i];
                if ((uint32_t) id >= n_batch()) {
                    throw std::runtime_error(format("invalid output id, %d does not fit in batch size of %u", id, n_batch()));
                }
                this->output_ids[id] = i;
            }

            this->n_outputs = n_outputs;
        }
    }

    // read logits
    {
        LLAMA_LOG_DEBUG("%s: - reading logits\n", __func__);

        uint64_t logits_size;
        io.read_to(&logits_size, sizeof(logits_size));

        if (this->logits_size < logits_size) {
            throw std::runtime_error("logits buffer too small");
        }

        if (logits_size) {