experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

In NamedGPUCombination.should_execute_combination, check if number_of_required_gpus > 0 before making a call to context.num_gpus(), since if number_of_required_gpus is zero, it does not even need to query the number of gpus available.

Browse Source 
This alleviates the requirement that collective ops needs to be configured at program startup for cases where gpu is not needed, because this code path is run when using combinations, which forces collective ops to be configured even before the test body is run.

PiperOrigin-RevId: 281685344
Change-Id: Ic53c3a50c95c512bd88cf685a6f50c1867b469b7
This commit is contained in:
Rick Chao 2019-11-20 23:39:46 -08:00 committed by TensorFlower Gardener
parent a2435de245
commit 915bef5c6a
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
tensorflow/python/distribute/combinations.py
View File

@ -79,7 +79,8 @@ class NamedGPUCombination(test_combinations.TestCombination):
if not number_of_required_gpus and GPUCombination.GPU_TEST:
return (False, "Test that doesn't require GPUs.")
elif context.num_gpus() < number_of_required_gpus:
elif (number_of_required_gpus > 0
and context.num_gpus() < number_of_required_gpus):
return (False, ("Only {} of {} required GPUs are available.".format(
context.num_gpus(), number_of_required_gpus)))
else: