I am using distributed training with MultiWorkerMirroredStrategy in Tensorflow 1.14 (currently in the progress of moving to Tensorflow 2.4), and I wanted to understand if there is a need to backport any changes from ScopedAllocatorOptimizer, and if so, which changes would be critical. Looking at
https://github.com/tensorflow/tensorflow/commits/master/tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc it seems like there are issues with control_dependencies for example, which makes me think that some issues I've been seeing with MultiWorkerMirroredStrategy might be due to my version of Tensorflow. Thank you!