extend

rwth-i6/returnn#300
tensorflow/tensorflow#9374
tensorflow/tensorflow#8021
tensorflow/tensorflow#8136

tf-gpu-mem-fraction.py

@@ -1,23 +1,35 @@
 #!/usr/bin/env python3
 
 """
-Via: https://github.com/rwth-i6/returnn/issues/300
+Via:
+https://github.com/rwth-i6/returnn/issues/300
+https://github.com/tensorflow/tensorflow/issues/9374
+https://github.com/tensorflow/tensorflow/issues/8021
+https://github.com/tensorflow/tensorflow/issues/8136
+
 Reported by Akshat Dewan <https://github.com/akshatdewan>.
 """
 
 import tensorflow as tf
 from tensorflow.python.client import device_lib
 import subprocess
 
+print("TF version:", tf.__version__)
 tf.compat.v1.disable_eager_execution()
 
 v = tf.compat.v1.get_variable("var", shape=(100, 100))
 result = v
 
+gpu_options = tf.compat.v1.GPUOptions(per_process_gpu_memory_fraction=0.1)
+pre_ops = {}
+pre_ops["device_count"] = {"GPU": 0}  # this will ignore the gpu_options for now...
+
+with tf.compat.v1.Session(config=tf.compat.v1.ConfigProto(gpu_options=gpu_options, **pre_ops)) as sess:
+  pass
+
 # The following line will cause the full GPU mem allocation.
 print(list(device_lib.list_local_devices()))
 
-gpu_options = tf.compat.v1.GPUOptions(per_process_gpu_memory_fraction=0.1)
 
 with tf.compat.v1.Session(config=tf.compat.v1.ConfigProto(gpu_options=gpu_options)) as sess:
   sess.run(tf.compat.v1.global_variables_initializer())