def get_computation_devices(
preferred_gpu_list: Optional[List[int]],
multi_gpu_flag: bool,
) -> List[Device]:
"""get the available computation devices (CPU & GPUs)
Get the computation devices for deep learning experiments with given
preferred list of GPU and flag for multi-GPU computation.
:param preferred_gpu_list: preferred list of GPUs represented with
integers starting from 0. For instances, [0, 2] represents th first and
the third GPUs. None or empty list of GPUS indicate the usage of CPU
:param multi_gpu_flag: boolean flag for multi-GPU training and testing
:return: list of integers representing the available devices (CPU & GPUs)
"""
# use CPU when GPUs are not preferred or not available
if (preferred_gpu_list is None) \
or (len(preferred_gpu_list) == 0) \
or (not torch.cuda.is_available()):
return [
Device('cpu'),
]
# else GPUs are preferred and available
# get all available GPU indexes
_available_gpu_list: List[int]
if getAvailable:
# by default, use GPU utility package with load and memory usage
# specification so that the 'available' GPUs are actually ready
# for deep learning runs (https://github.com/anderskm/gputil)
_available_gpu_list = getAvailable(
limit=_MAX_NUM_GPUS,
maxLoad=_MAX_GPU_LOAD,
maxMemory=_MAX_GPU_MEM_USED,
)
else:
# assume all GPUs are good to use without GPUtil package
_available_gpu_list = list(range(torch.cuda.device_count()))
_warning_msg = \
f'GPUtil (https://github.com/anderskm/gputil) not installed.' \
f'Assuming all GPUs ({_available_gpu_list}) are available ' \
f'and ready for training ... '
_LOGGER.warning(_warning_msg)
# get the overlap between the preferred and the available GPUs
_gpus = \
[_g for _g in _available_gpu_list if _g in preferred_gpu_list]
# use CPU if there is no preferred GPUs that are available
if len(_gpus) == 0:
return [
Device('cpu'),
]
# otherwise return one or all GPUs depending on the multi-GPU flag
return [Device(f'cuda:{_g}') for _g in _gpus] \
if multi_gpu_flag else [Device(f'cuda:{_gpus[0]}'), ]
def auto_gpus_select(maxGpuNum):
from GPUtil import getAvailable
from time import sleep
while True:
gpu_idxs = getAvailable(order='memory',
limit=maxGpuNum,
maxLoad=0.2,
maxMemory=0.2,
includeNan=False)
if len(gpu_idxs):
break
else:
print("No available GPU! Wait for other users to exit...")
sleep(10)
return gpu_idxs
def get_gpus():
return getAvailable(order="random", limit=8, maxLoad=10**-6, maxMemory=10**-2)
def get_computation_devices(
preferred_gpu_list: Optional[Union[List[int], str]],
multi_gpu_flag: bool,
) -> List[Device]:
"""get the available computation devices (CPU & GPUs)
Get the computation devices for deep learning experiments with given
preferred list of GPU and flag for multi-GPU computation.
:param preferred_gpu_list: preferred list of GPUs represented with
integers starting from 0; e.g. [0, 2] represents th first and
the third GPUs; none or empty list of GPUS indicate the usage of CPU;
and 'all' indicates all GPUs are preferred
:type preferred_gpu_list: Optional[Union[List[int], str]]
:param multi_gpu_flag: boolean flag for multi-GPU training and testing
:type multi_gpu_flag: bool
:return: list of available computation devices (CPU & GPUs) visible to
PyTorch (not hardware indices from PCIe)
:rtype: List[Device]
"""
# use CPU when GPUs are not preferred or not available
if (preferred_gpu_list is None) \
or (len(preferred_gpu_list) == 0) \
or (not torch.cuda.is_available()):
return [Device('cpu'), ]
# else GPUs are preferred and available
# get all available GPU indexes
_available_gpu_list: List[int]
if getAvailable:
# by default, use GPU utility package with load and memory usage
# specification so that the 'available' GPUs are actually ready
# for deep learning runs (https://github.com/anderskm/gputil)
_available_gpu_list = getAvailable(
limit=_MAX_NUM_GPUS,
maxLoad=_MAX_GPU_LOAD,
maxMemory=_MAX_GPU_MEM_USED,
)
else:
# assume all GPUs are good to use without GPUtil package
_available_gpu_list = list(range(torch.cuda.device_count()))
_warning_msg = \
f'GPUtil (https://github.com/anderskm/gputil) not installed.' \
f'Assuming all GPUs ({_available_gpu_list}) are available ' \
f'and ready for training ... '
_LOGGER.warning(_warning_msg)
# if CUDA_VISIBLE_DEVICES is set as an environment variable, then make
# sure that the list of available GPUs are visible, and re-index them in
# the way that PyTorch can access. For example:
# CUDA_VISIBLE_DEVICES = [4, 5, 6, 7]
# GPUs not in use = [1, 4, 7]
# available GPUs = [4, 7]
# available GPUs for PyTorch = [0, 3]
if 'CUDA_VISIBLE_DEVICES' in os.environ:
_available_gpu_list_: List[int] = []
_cuda_visible_devices: List[int] = \
[int(_i) for _i in os.environ['CUDA_VISIBLE_DEVICES'].split(',')]
for __available_gpu_id in _available_gpu_list:
if __available_gpu_id in _cuda_visible_devices:
_available_gpu_list_.append(
_cuda_visible_devices.index(__available_gpu_id))
_available_gpu_list = _available_gpu_list_
# double check to make sure that all GPUs are accessible for PyTorch
_available_gpu_list_: List[int] = []
for __available_gpu_id in _available_gpu_list:
try:
torch.cuda.get_device_properties(__available_gpu_id)
except Exception:
_warning_msg = \
f'CUDA device {__available_gpu_id}, despite in the range ' \
f'of PyTorch GPU count, is not available.'
_LOGGER.warning(_warning_msg)
print(_warning_msg)
else:
_available_gpu_list_.append(__available_gpu_id)
_available_gpu_list = _available_gpu_list_
# get the overlap between the preferred and the available GPUs
if isinstance(preferred_gpu_list, str) and preferred_gpu_list == 'all':
_gpus = _available_gpu_list
elif isinstance(preferred_gpu_list, list):
_gpus = list(set(_available_gpu_list).intersection(preferred_gpu_list))
else:
_error_msg = \
f'Unknown parameter {preferred_gpu_list} for the list of ' \
f'preferred GPUs. Must be either \"all\" or list of integers.'
raise ValueError(_error_msg)
# use CPU if there is no preferred GPUs that are available
if len(_gpus) == 0:
return [Device('cpu'), ]
# otherwise return one or all GPUs depending on the multi-GPU flag
return [Device(f'cuda:{_g}') for _g in _gpus] \
if multi_gpu_flag else [Device(f'cuda:{_gpus[0]}'), ]
def get_gpus():
return [
str(idx)
for idx in getAvailable(limit=8, maxLoad=10**-1, maxMemory=10**-1)
]