geowatch.utils.util_nvidia module

Ported from netharn.device, previously called gpu_infos

exception geowatch.utils.util_nvidia.NvidiaSMIError

Bases: Exception

geowatch.utils.util_nvidia.nvidia_smi(ignore_environ=False)

Run nvidia-smi and parse output

Parameters:

• new_mode – internal argument that changes the underlying implementation

• ignore_environ (bool) – if True respects CUDA_VISIBLE_DEVICES environment variable, otherwise returns data corresponding to physical GPU indexes. Defaults to False.

Returns:

info about each nvidia GPU indexed by gpu number

Return type:

dict

Note

Not gaurenteed to work if CUDA is not installed.

Warning

if nvidia-smi is not found

Example

>>> # xdoctest: +REQUIRES(env:HAS_CUDA)
>>> from geowatch.utils.util_nvidia import *  # NOQA
>>> gpus = nvidia_smi()
>>> # xdoctest: +IGNORE_WANT
>>> import torch
>>> print('gpus = {}'.format(ub.repr2(gpus, nl=4)))
>>> assert len(gpus) == torch.cuda.device_count()
gpus = {
    0: {
        'gpu_uuid': 'GPU-348ebe36-252b-46fa-8a97-477ae331f6f4',
        'index': '0',
        'mem_avail': 10013.0,
        'mem_total': 11170.0,
        'mem_used': 1157.0,
        'memory.free': '10013 MiB',
        'memory.total': '11170 MiB',
        'memory.used': '1157 MiB',
        'name': 'GeForce GTX 1080 Ti',
        'num': 0,
        'num_compute_procs': 1,
        'procs': [
            {
                'gpu_num': 0,
                'gpu_uuid': 'GPU-348ebe36-252b-46fa-8a97-477ae331f6f4',
                'name': '/usr/bin/python',
                'pid': '19912',
                'type': 'C',
                'used_memory': '567 MiB',
            },
        ],
    },
}