Device Handling in Φ_ML

  •   🌐 Φ_ML   •   📖 Documentation   •   🔗 API   •   ▶ Videos   •   Examples

This notebook is work in progress. It will explain

• Selecting a device CPU / GPU
• to_device()

%%capture
!pip install phiml

from phiml import math, backend

Compute Devices

Φ_ML abstracts ComputeDevices, such as CPUs, GPUs and TPUs. You can obtain a list of available devices using Backend.list_devices()

BACKEND = math.use('torch')
BACKEND.list_devices('CPU')

[torch device 'CPU' (CPU 'cpu') | 15990 MB | 4 processors | ]

Compute devices are bound to the computing library that uses it, i.e. the CPU returned by PyTorch is not the same object as the CPU returned by NumPy.

from phiml.backend import NUMPY
NUMPY.list_devices('CPU')

[numpy device 'CPU' (CPU 'CPU') | 15990 MB | 4 processors | ]

Setting the Default Device

We can set the default device per backend by reference or type.

BACKEND.set_default_device(BACKEND.list_devices()[0])

True

BACKEND.set_default_device('GPU')

/opt/hostedtoolcache/Python/3.8.18/x64/lib/python3.8/site-packages/phiml/backend/_backend.py:261: RuntimeWarning: torch: Cannot select 'GPU' because no device of this type is available.
  warnings.warn(f"{self.name}: Cannot select '{device}' because no device of this type is available.", RuntimeWarning)

False

Tensors created by that backend will be allocated on that device from now on. Already allocated tensors are left untouched.

Combining tensors allocated on different devices may lead to errors!

Moving Tensors to a Different Device

You can move any tensor to a different compute device

tensor = math.zeros()
tensor.device

torch device 'CPU' (CPU 'cpu') | 15990 MB | 4 processors | 

math.to_device(tensor, 'CPU')

0.0

math.to_device() also supports pytrees and data classes that contain tensors.

Further Reading

Φ_ML also supports moving tensors to different backend libraries without copying them.

🌐 Φ_ML   •   📖 Documentation   •   🔗 API   •   ▶ Videos   •   Examples