Python pynvml.nvmlDeviceGetCount方法代码示例

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def getFreeId():
    import pynvml 

    pynvml.nvmlInit()
    def getFreeRatio(id):
        handle = pynvml.nvmlDeviceGetHandleByIndex(id)
        use = pynvml.nvmlDeviceGetUtilizationRates(handle)
        ratio = 0.5*(float(use.gpu+float(use.memory)))
        return ratio

    deviceCount = pynvml.nvmlDeviceGetCount()
    available = []
    for i in range(deviceCount):
        if getFreeRatio(i)<70:
            available.append(i)
    gpus = ''
    for g in available:
        gpus = gpus+str(g)+','
    gpus = gpus[:-1]
    return gpus 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def getGPUUsage():
    try:
        pynvml.nvmlInit()
        count = pynvml.nvmlDeviceGetCount()
        if count == 0:
            return None

        result = {"driver": pynvml.nvmlSystemGetDriverVersion(),
                  "gpu_count": int(count)
                  }
        i = 0
        gpuData = []
        while i<count:
            handle = pynvml.nvmlDeviceGetHandleByIndex(i)
            mem = pynvml.nvmlDeviceGetMemoryInfo(handle)
            gpuData.append({"device_num": i, "name": pynvml.nvmlDeviceGetName(handle), "total": round(float(mem.total)/1000000000, 2), "used": round(float(mem.used)/1000000000, 2)})
            i = i+1

        result["devices"] = jsonpickle.encode(gpuData, unpicklable=False)
    except Exception as e:
        result = {"driver": "No GPU!", "gpu_count": 0, "devices": []}

    return result 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def getFreeId():
    import pynvml

    pynvml.nvmlInit()

    def getFreeRatio(id):
        handle = pynvml.nvmlDeviceGetHandleByIndex(id)
        use = pynvml.nvmlDeviceGetUtilizationRates(handle)
        ratio = 0.5 * (float(use.gpu + float(use.memory)))
        return ratio

    deviceCount = pynvml.nvmlDeviceGetCount()
    available = []
    for i in range(deviceCount):
        if getFreeRatio(i) < 70:
            available.append(i)
    gpus = ''
    for g in available:
        gpus = gpus + str(g) + ','
    gpus = gpus[:-1]
    return gpus 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def get_appropriate_cuda(task_scale='s'):
    if task_scale not in {'s','m','l'}:
        logger.info('task scale wrong!')
        exit(2)
    import pynvml
    pynvml.nvmlInit()
    total_cuda_num = pynvml.nvmlDeviceGetCount()
    for i in range(total_cuda_num):
        logger.info(i)
        handle = pynvml.nvmlDeviceGetHandleByIndex(i)  # 这里的0是GPU id
        memInfo = pynvml.nvmlDeviceGetMemoryInfo(handle)
        utilizationInfo = pynvml.nvmlDeviceGetUtilizationRates(handle)
        logger.info(i, 'mem:', memInfo.used / memInfo.total, 'util:',utilizationInfo.gpu)
        if memInfo.used / memInfo.total < 0.15 and utilizationInfo.gpu <0.2:
            logger.info(i,memInfo.used / memInfo.total)
            return 'cuda:'+str(i)

    if task_scale=='s':
        max_memory=2000
    elif task_scale=='m':
        max_memory=6000
    else:
        max_memory = 9000

    max_id = -1
    for i in range(total_cuda_num):
        handle = pynvml.nvmlDeviceGetHandleByIndex(0)  # 这里的0是GPU id
        memInfo = pynvml.nvmlDeviceGetMemoryInfo(handle)
        utilizationInfo = pynvml.nvmlDeviceGetUtilizationRates(handle)
        if max_memory < memInfo.free:
            max_memory = memInfo.free
            max_id = i

    if id == -1:
        logger.info('no appropriate gpu, wait!')
        exit(2)

    return 'cuda:'+str(max_id)

        # if memInfo.used / memInfo.total < 0.5:
        #     return 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def device_count_for():
    try:
        return pynvml.nvmlDeviceGetCount()
    except pynvml.NVMlError:
        return None 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def record_measurements_to(async_reporting_func, polling_interval=1):
    try:
        deviceCount = pynvml.nvmlDeviceGetCount()
        while True:
            measurement = await aggregate_measurements(deviceCount)
            await async_reporting_func(measurement)
            await asyncio.sleep(polling_interval)
    except CancelledError:
        _logger().info("Logging cancelled") 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def _get_device_count(self):
        """ Detect the number of GPUs attached to the system and allocate to
        :attr:`_device_count`. """
        if self._is_plaidml:
            self._device_count = self._plaid.device_count
        elif IS_MACOS:
            self._device_count = pynvx.cudaDeviceGetCount(ignore=True)
        else:
            try:
                self._device_count = pynvml.nvmlDeviceGetCount()
            except pynvml.NVMLError:
                self._device_count = 0
        self._log("debug", "GPU Device count: {}".format(self._device_count)) 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def get_gpu_metrics() -> List:
    try:
        pynvml.nvmlInit()
        device_count = pynvml.nvmlDeviceGetCount()
        results = []

        for i in range(device_count):
            handle = pynvml.nvmlDeviceGetHandleByIndex(i)
            results += metrics_dict_to_list(query_gpu(handle))
        return results
    except pynvml.NVMLError:
        logger.debug("Failed to collect gpu resources", exc_info=True)
        return [] 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def _find_gpu(self):
        device_count = pynvml.nvmlDeviceGetCount()
        for i in range(device_count):
            handle = pynvml.nvmlDeviceGetHandleByIndex(i)
            gpu_processes = pynvml.nvmlDeviceGetComputeRunningProcesses(handle)
            for gpu_process in gpu_processes:
                if gpu_process.pid == self.pid:
                    self.gpu = handle
        
        self.accounting_enabled = pynvml.nvmlDeviceGetAccountingMode(self.gpu) == pynvml.NVML_FEATURE_ENABLED
        
        # Clear accounting statistics (requires root privileges)
        #pynvml.nvmlDeviceSetAccountingMode(self.gpu, pynvml.NVML_FEATURE_DISABLED)
        #pynvml.nvmlDeviceSetAccountingMode(self.gpu, pynvml.NVML_FEATURE_ENABLED) 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetCount [as 别名]
def _crawl_in_system(self):
        '''
        nvidia-smi returns following: MEMORY, UTILIZATION, ECC, TEMPERATURE,
        POWER, CLOCK, COMPUTE, PIDS, PERFORMANCE, SUPPORTED_CLOCKS,
        PAGE_RETIREMENT, ACCOUNTING

        currently, following are requested based on dlaas requirements:
            utilization.gpu, utilization.memory,
            memory.total, memory.free, memory.used
        nvidia-smi --query-gpu=utilization.gpu,utilization.memory,\
            memory.total,memory.free,memory.used --format=csv,noheader,nounits
        '''

        if self._init_nvml() == -1:
            return

        self.inspect_arr = exec_dockerps()

        num_gpus = pynvml.nvmlDeviceGetCount()

        for gpuid in range(0, num_gpus):
            gpuhandle = pynvml.nvmlDeviceGetHandleByIndex(gpuid)
            temperature = pynvml.nvmlDeviceGetTemperature(
                gpuhandle, pynvml.NVML_TEMPERATURE_GPU)
            memory = pynvml.nvmlDeviceGetMemoryInfo(gpuhandle)
            mem_total = memory.total / 1024 / 1024
            mem_used = memory.used / 1024 / 1024
            mem_free = memory.free / 1024 / 1024
            power_draw = pynvml.nvmlDeviceGetPowerUsage(gpuhandle) / 1000
            power_limit = pynvml.nvmlDeviceGetEnforcedPowerLimit(
                gpuhandle) / 1000
            util = pynvml.nvmlDeviceGetUtilizationRates(gpuhandle)
            util_gpu = util.gpu
            util_mem = util.memory
            entry = {
                'utilization': {'gpu': util_gpu, 'memory': util_mem},
                'memory': {'total': mem_total, 'free': mem_free,
                           'used': mem_used},
                'temperature': temperature,
                'power': {'draw': power_draw, 'limit': power_limit}
            }
            key = self._get_feature_key(gpuhandle, gpuid)
            if gpuid == num_gpus - 1:
                self._shutdown_nvml()

            yield (key, entry, 'gpu')

        return