Felipe97/llama-cpp-compiled

1// ref: https://github.com/ggml-org/llama.cpp/issues/259372// only works reliably when run with a large model that occupies 3GB+ of wired memory3// thus, this test is not run by default4// example model to run with: google/gemma-4-E4B-it-qat-q4_0-gguf5 6#include "llama.h"7#include "common.h"8 9#include <cstdint>10#include <mach/mach.h>11#include <mach/mach_host.h>12#include <unistd.h>13 14static uint64_t wired_memory() {15    vm_statistics64_data_t vmstat;16    mach_msg_type_number_t count = HOST_VM_INFO64_COUNT;17    if (host_statistics64(mach_host_self(), HOST_VM_INFO64, (host_info64_t)&vmstat, &count) != KERN_SUCCESS) {18        return UINT64_MAX;19    }20    return static_cast<uint64_t>(vmstat.wire_count) * vm_kernel_page_size;21}22 23int main(int argc, char ** argv) {24    auto * model_path = common_get_model_or_exit(argc, argv);25 26    llama_backend_init();27 28    const uint64_t wired_initial = wired_memory();29 30    llama_model_params params = llama_model_default_params();31    params.load_mode = LLAMA_LOAD_MODE_NONE;32    struct llama_model* model = llama_model_load_from_file(model_path, params);33 34    const uint64_t wired_loaded = wired_memory();35    const uint64_t wired_delta = wired_loaded - wired_initial;36    // system memory fluctuates, so we need to allocate enough to reliably detect the release37    GGML_ASSERT(wired_delta > 2'000'000'000); // 2GB38 39    llama_model_free(model);40 41    const uint64_t t_start_ms = ggml_time_ms();42 43    // expect most of the allocated memory to be released within 10 seconds44    // we allow for some tolerance due to system-wide memory fluctuations45    while (wired_memory() > wired_loaded - 0.75 * wired_delta) {46        GGML_ASSERT(ggml_time_ms() - t_start_ms < 10'000);47        usleep(100'000); // 100ms48    }49 50    llama_backend_free();51 52    return 0;53}54