Source code for ignite.metrics.gpu_info

# -*- coding: utf-8 -*-
import warnings
from typing import Any, Dict, List, Tuple, Union

import torch

from ignite.engine import Engine, EventEnum, Events
from ignite.metrics.metric import Metric


class GpuInfo(Metric):
    """Provides GPU information: a) used memory percentage, b) gpu utilization percentage values as Metric
    on each iterations. This metric requires `pynvml <https://pypi.org/project/pynvml/>`_ package of version `<12`.

    .. Note ::

        In case if gpu utilization reports "N/A" on a given GPU, corresponding metric value is not set.

    Examples:
        .. code-block:: python

            # Default GPU measurements
            GpuInfo().attach(trainer, name='gpu')  # metric names are 'gpu:X mem(%)', 'gpu:X util(%)'

            # Logging with TQDM
            ProgressBar(persist=True).attach(trainer, metric_names=['gpu:0 mem(%)', 'gpu:0 util(%)'])
            # Progress bar will looks like
            # Epoch [2/10]: [12/24]  50%|█████      , gpu:0 mem(%)=79, gpu:0 util(%)=59 [00:17<1:23]

            # Logging with Tensorboard
            tb_logger.attach(trainer,
                             log_handler=OutputHandler(tag="training", metric_names='all'),
                             event_name=Events.ITERATION_COMPLETED)
    """

    def __init__(self) -> None:
        try:
            from pynvml.smi import nvidia_smi
        except ImportError:
            raise ModuleNotFoundError(
                "This contrib module requires pynvml to be installed. "
                "Please install it with command: \n pip install 'pynvml<12'"
            )
            # Let's check available devices
        if not torch.cuda.is_available():
            raise RuntimeError("This contrib module requires available GPU")

        # Let it fail if no libnvidia drivers or NMVL library found
        self.nvsmi = nvidia_smi.getInstance()
        super(GpuInfo, self).__init__()

    def reset(self) -> None:
        pass

    def update(self, output: Tuple[torch.Tensor, torch.Tensor]) -> None:
        pass

    def compute(self) -> List[Dict[str, Any]]:
        data: Dict[str, List[Dict[str, Any]]] = self.nvsmi.DeviceQuery("memory.used, memory.total, utilization.gpu")
        if len(data) == 0 or ("gpu" not in data):
            warnings.warn("No GPU information available")
            return []
        return data["gpu"]

    def completed(self, engine: Engine, name: str) -> None:
        data = self.compute()
        if len(data) < 1:
            warnings.warn("No GPU information available")
            return

        for i, data_by_rank in enumerate(data):
            mem_name = f"{name}:{i} mem(%)"

            if "fb_memory_usage" not in data_by_rank:
                warnings.warn(f"No GPU memory usage information available in {data_by_rank}")
                continue
            mem_report = data_by_rank["fb_memory_usage"]
            if not ("used" in mem_report and "total" in mem_report):
                warnings.warn(
                    "GPU memory usage information does not provide used/total "
                    f"memory consumption information in {mem_report}"
                )
                continue

            engine.state.metrics[mem_name] = int(mem_report["used"] * 100.0 / mem_report["total"])

        for i, data_by_rank in enumerate(data):
            util_name = f"{name}:{i} util(%)"
            if "utilization" not in data_by_rank:
                warnings.warn(f"No GPU utilization information available in {data_by_rank}")
                continue
            util_report = data_by_rank["utilization"]
            if not ("gpu_util" in util_report):
                warnings.warn(f"GPU utilization information does not provide 'gpu_util' information in {util_report}")
                continue
            try:
                engine.state.metrics[util_name] = int(util_report["gpu_util"])
            except ValueError:
                # Do not set GPU utilization information
                pass

    # TODO: see issue https://github.com/pytorch/ignite/issues/1405
    def attach(  # type: ignore
        self, engine: Engine, name: str = "gpu", event_name: Union[str, EventEnum] = Events.ITERATION_COMPLETED
    ) -> None:
        engine.add_event_handler(event_name, self.completed, name)