Function main

subprojects/llama.cpp/tests/test-thread-safety.cpp:16–164

#include "sampling.h"

int main(int argc, char ** argv) {
    common_params params;

    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
        return 1;
    }

    common_init();

    llama_backend_init();
    llama_numa_init(params.numa);

    LOG_INF("%s\n", common_params_get_system_info(params).c_str());

    //llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
    //    if (level == GGML_LOG_LEVEL_ERROR) {
    //        common_log_add(common_log_main(), level, "%s", text);
    //    }
    //}, NULL);

    auto cparams = common_context_params_to_llama(params);

    // each context has a single sequence
    cparams.n_seq_max = 1;

    int dev_count = ggml_backend_dev_count();
    std::vector<std::array<ggml_backend_dev_t, 2>> gpus;
    for (int i = 0; i < dev_count; ++i) {
        auto * dev = ggml_backend_dev_get(i);
        if (dev && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
            gpus.push_back({dev, nullptr});
        }
    }
    const int gpu_dev_count = (int)gpus.size();
    const int num_models = gpu_dev_count + 1 + 1; // GPUs + 1 CPU model + 1 layer split
    //const int num_models = std::max(1, gpu_dev_count);
    const int num_contexts = std::max(1, params.n_parallel);

    std::vector<llama_model_ptr> models;
    std::vector<std::thread> threads;
    std::atomic<bool> failed = false;

    for (int m = 0; m < num_models; ++m) {
        auto mparams = common_model_params_to_llama(params);

        if (m < gpu_dev_count) {
            mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
            mparams.devices = gpus[m].data();
        } else if (m == gpu_dev_count) {
            mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
            mparams.main_gpu = -1; // CPU model
        } else {
            mparams.split_mode = LLAMA_SPLIT_MODE_LAYER;
        }

        llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
        if (model == NULL) {
            LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());