Python pynvml.nvmlDeviceGetName方法代码示例

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetName [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 nvmlDeviceGetName [as 别名]
def gpu_info(self):
        # pip install nvidia-ml-py3
        if len(self.gpu_ids) >= 0 and torch.cuda.is_available():
            try:
                import pynvml
                pynvml.nvmlInit()
                self.config_dic['gpu_driver_version'] = pynvml.nvmlSystemGetDriverVersion()
                for gpu_id in self.gpu_ids:
                    handle = pynvml.nvmlDeviceGetHandleByIndex(gpu_id)
                    gpu_id_name = "gpu%s" % gpu_id
                    mem_info = pynvml.nvmlDeviceGetMemoryInfo(handle)
                    gpu_utilize = pynvml.nvmlDeviceGetUtilizationRates(handle)
                    self.config_dic['%s_device_name' % gpu_id_name] = pynvml.nvmlDeviceGetName(handle)
                    self.config_dic['%s_mem_total' % gpu_id_name] = gpu_mem_total = round(mem_info.total / 1024 ** 3, 2)
                    self.config_dic['%s_mem_used' % gpu_id_name] = gpu_mem_used = round(mem_info.used / 1024 ** 3, 2)
                    # self.config_dic['%s_mem_free' % gpu_id_name] = gpu_mem_free = mem_info.free // 1024 ** 2
                    self.config_dic['%s_mem_percent' % gpu_id_name] = round((gpu_mem_used / gpu_mem_total) * 100, 1)
                    self._set_dict_smooth('%s_utilize_gpu' % gpu_id_name, gpu_utilize.gpu, 0.8)
                    # self.config_dic['%s_utilize_gpu' % gpu_id_name] = gpu_utilize.gpu
                    # self.config_dic['%s_utilize_memory' % gpu_id_name] = gpu_utilize.memory

                pynvml.nvmlShutdown()
            except Exception as e:
                print(e) 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetName [as 别名]
def _get_devices(self):
        """ Obtain the name of the installed devices. The quality of this information depends on
        the backend and OS being used, but it should be sufficient for identifying cards.

        Returns
        -------
        list
            List of device names for connected GPUs as corresponding to the values in
            :attr:`_handles`
        """
        self._initialize()
        if self._device_count == 0:
            names = list()
        if self._is_plaidml:
            names = self._plaid.names
        elif IS_MACOS:
            names = [pynvx.cudaGetName(handle, ignore=True)
                     for handle in self._handles]
        else:
            names = [pynvml.nvmlDeviceGetName(handle).decode("utf-8")
                     for handle in self._handles]
        self._log("debug", "GPU Devices: {}".format(names))
        return names 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetName [as 别名]
def device_name_for(device_handle):
    """Get GPU device name"""
    try:
        return nativestr(pynvml.nvmlDeviceGetName(device_handle))
    except pynvml.NVMlError:
        return "NVIDIA" 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetName [as 别名]
def measure_cpu_gpu_instant_load():
    # Get current cpu gpu load, as
    # load = [rank, cpu_load, nvidia_device_id, gpu_load]
    # result_arr: [load, load, ...]

    if cg_load_backend_ok:
        global gpu_a_load
        global gpu_m_count
        global p_handler

        cpu_load = p_handler.cpu_percent()
        gpu_m_count += 1
        try:
            comm = current_communicator()
            if comm:
                index = comm.local_rank
            elif 'cuda' in str(nn.get_current_context().backend):
                index = 0
            else:
                raise Exception
            handler = pynvml.nvmlDeviceGetHandleByIndex(index)
            gpu_load = [
                [index, pynvml.nvmlDeviceGetUtilizationRates(handler).gpu]]

            if index in gpu_a_load.keys():
                gpu_a_load[index]['name'] = pynvml.nvmlDeviceGetName(
                    handler).decode("utf-8")
                o_load = gpu_a_load[index]['load']
                n_load = gpu_load[0][1]
                gpu_a_load[index]['load'] = (
                    (gpu_m_count - 1) * o_load + n_load) / gpu_m_count
            else:
                gpu_a_load[index] = {
                    'name': pynvml.nvmlDeviceGetName(handler).decode("utf-8"),
                    'load': gpu_load[0][1]
                }

        except Exception:
            gpu_load = [[-1, -1]]

        callback.update_status(
            ('cpu_gpu_load', collect_and_shape_result(cpu_load, gpu_load))) 

# 需要导入模块: import pynvml [as 别名]
# 或者: from pynvml import nvmlDeviceGetName [as 别名]
def _log_statistics(self, elapsed_steps, elapsed_time, global_step):
        """Collect and store all summary values.

        Arguments:
            elapsed_steps (int):
                The number of steps between the current trigger event and the last one.
            elapsed_time (float):
                The number of seconds between the current trigger event and the last one.
            global_step (tf.Tensor):
                Global step tensor.
        """

        # Write summary for tensorboard.
        if self._summary_writer is not None:
            summary_list = list()
            # Add only summaries.
            for gpu_id in self._gpu_statistics.keys():
                for statistic in self._gpu_statistics[gpu_id].keys():
                    # only add them if they are requested for logging.
                    if statistic in self._statistics_to_log:
                        values = self._gpu_statistics[gpu_id][statistic]
                        # Only Calculate and write average if there is data available.
                        if values:
                            avg_value = sum(values) / len(values)
                            avg_summary = Summary.Value(tag='{}/{}:{}'
                                                        .format(self._group_tag, gpu_id, statistic),
                                                        simple_value=avg_value)
                            summary_list.append(avg_summary)

            # Write all statistics as simple scalar summaries.
            summary = Summary(value=summary_list)
            self._summary_writer.add_summary(summary, global_step)

        # Log summaries to the logging stream.
        if not self._suppress_stdout:
            for gpu_id in self._gpu_statistics.keys():
                # Acquire a GPU device handle.
                handle = nvml.nvmlDeviceGetHandleByIndex(gpu_id)

                # Query the device name.
                name = nvml.nvmlDeviceGetName(handle).decode('utf-8')

                for statistic in self._gpu_statistics[gpu_id].keys():
                    # Log utilization information with INFO level.
                    logging.debug("%s: %s", name, '{}: {}'
                                  .format(statistic, self._gpu_statistics[gpu_id][statistic]))


# The following code has been inspired by <https://stackoverflow.com/a/45681782>: