How to list visible Tensorflow devices without allocation?

Question

One can list the currently visible devices to Tensorflow with:

from tensorflow.python.client import device_lib
LOCAL_DEVICES = device_lib.list_local_devices()

(see: How to get current available GPUs in tensorflow?)

However, this seems to initialize and allocate these GPUs. How can we list the devices without allocating them?

What gives you the indication that the listed GPUs have been allocated after call to device_lib? — brown.2179, Oct 14 '17 at 20:00

score 2 · Accepted Answer · answered Oct 13 '17 at 18:16

For GPU devices, the TensorFlow list_local_devices() function is a wrapper around the cuDeviceGetCount(), cuDeviceGet(), and cuDeviceGetProperties() functions in the CUDA API. (In addition, there is typically one "CPU device" that uses all of the cores in the local system.)

The simplest solution might be to call these CUDA APIs directly. This script shows how to do this from Python using ctypes.

How to list visible Tensorflow devices without allocation?

1 Answers1