cellmap_flow.utils.device
=========================

.. py:module:: cellmap_flow.utils.device

.. autoapi-nested-parse::

   The device torch work runs on, wherever cellmap-flow picks one.



Functions
---------

.. autoapisummary::

   cellmap_flow.utils.device.best_device


Module Contents
---------------

.. py:function:: best_device()

   CUDA when there is one, else Apple's GPU (MPS), else the CPU.

   Every choice here was CUDA or the CPU, so a Mac served and trained its
   models on the CPU, though PyTorch runs on its GPU and Cellpose and
   bioimageio.core already used it on their own: a Cellpose finetune took
   about 33 s a batch of eight tiles there, two days for 20 epochs.