Method state_write_data

smallthinker/src/llama-context.cpp:1754–1830

}

size_t llama_context::state_write_data(llama_io_write_i & io) {
    LLAMA_LOG_DEBUG("%s: writing state\n", __func__);

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

        const std::string arch_str = llm_arch_name(model.arch);
        io.write_string(arch_str);
        // TODO: add more model-specific info which should prevent loading the session file if not identical
    }

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

        const auto n_outputs    = this->n_outputs;
        const auto & output_ids = this->output_ids;

        std::vector<int32_t> w_output_pos;

        GGML_ASSERT(n_outputs <= n_outputs_max);

        w_output_pos.resize(n_outputs);

        // build a more compact representation of the output ids
        for (size_t i = 0; i < n_batch(); ++i) {
            // map an output id to a position in the batch
            int32_t pos = output_ids[i];
            if (pos >= 0) {
                GGML_ASSERT(pos < n_outputs);
                w_output_pos[pos] = i;
            }
        }

        io.write(&n_outputs, sizeof(n_outputs));

        if (n_outputs) {
            io.write(w_output_pos.data(), n_outputs * sizeof(int32_t));
        }
    }

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

        const uint64_t logits_size = std::min((uint64_t) this->logits_size, (uint64_t) n_outputs * model.vocab.n_tokens());

        io.write(&logits_size, sizeof(logits_size));

        if (logits_size) {
            io.write(logits, logits_size * sizeof(float));
        }
    }

    // write embeddings
    {
        LLAMA_LOG_DEBUG("%s: - writing embeddings\n", __func__);