Method state_seq_load_file

smallthinker/src/llama-context.cpp:1691–1732

}

size_t llama_context::state_seq_load_file(llama_seq_id seq_id, const char * filepath, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
    llama_file file(filepath, "rb");

    // version checks
    {
        const uint32_t magic   = file.read_u32();
        const uint32_t version = file.read_u32();

        if (magic != LLAMA_STATE_SEQ_MAGIC || version != LLAMA_STATE_SEQ_VERSION) {
            LLAMA_LOG_ERROR("%s: unknown (magic, version) for sequence state file: %08x, %08x\n", __func__, magic, version);
            return 0;
        }
    }

    // load the prompt
    {
        const uint32_t n_token_count = file.read_u32();

        if (n_token_count > n_token_capacity) {
            LLAMA_LOG_ERROR("%s: token count in sequence state file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
            return 0;
        }

        file.read_raw(tokens_out, sizeof(llama_token) * n_token_count);
        *n_token_count_out = n_token_count;
    }

    // restore the context state
    {
        const size_t state_size = file.size() - file.tell();
        llama_io_read_file io(&file);
        const size_t nread = state_seq_read_data(io, seq_id);
        if (!nread) {
            LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__);
            return 0;
        }
        GGML_ASSERT(nread <= state_size);
        GGML_ASSERT(nread + sizeof(uint32_t) * 3 + sizeof(llama_token) * *n_token_count_out == file.tell());
    }

    return file.tell();
}

size_t llama_context::state_seq_save_file(llama_seq_id seq_id, const char * filepath, const llama_token * tokens, size_t n_token_count) {
    llama_file file(filepath, "wb");