#include "llama_runtime.h"

#include "llama.h"

#include <clocale>

#include <cstdio>

#include <mutex>

#include <thread>

namespace vox::llama {

namespace {

void quiet_llama_log(ggml_log_level level, const char * text, void *) {

if (level == GGML_LOG_LEVEL_ERROR && text != nullptr) {

fputs(text, stderr);

}

}

std::mutex & runtime_mutex() {

static std::mutex value;

return value;

}

int & runtime_ref_count() {

static int value = 0;

return value;

}

} // namespace

LlamaRuntime::LlamaRuntime() {

std::lock_guard<std::mutex> lock(runtime_mutex());

if (runtime_ref_count() == 0) {

std::setlocale(LC_NUMERIC, "C");

llama_log_set(quiet_llama_log, nullptr);

llama_backend_init();

ggml_backend_load_all();

}

++runtime_ref_count();

}

LlamaRuntime::~LlamaRuntime() {

std::lock_guard<std::mutex> lock(runtime_mutex());

--runtime_ref_count();

if (runtime_ref_count() == 0) {

llama_backend_free();

}

}

int default_thread_count() {

const unsigned int count = std::thread::hardware_concurrency();

return count == 0 ? 4 : static_cast<int>(count);

}

} // namespace vox::llama