Show HN: Initial release of the distributed TensorFlow runtime
github.com
github.com
It's also hilarious that using distributed TensorFlow in Keras is a trivial change[1]. I honestly think high level libraries with transparent backends will be the future in the same way that cross compilation is the rule not the exception for compilers.
What's most impressive to me is that Google are incrementally open sourcing large parts of their internal infrastructure. Protocol Buffers, gRPC, Kubernetes, TensorFlow ... Given how cross dependent many of their internal dependencies likely are, extricating them is a major achievement.
Most importantly, it seems the more that's extricated, the easier it is to release more.
Can we now pretty please have it easier to build on systems commonly used for HPC?
Bazel won't compile on glibc that's used by RHEL6/CentOS6 or SLE11. Even if you do get that going, TensorFlow has an odd requirement for a very specific beta version of Protocol Buffers:
REQUIRED_PACKAGES = [ 'numpy >= %s' % numpy_version, 'six >= 1.10.0', 'protobuf == 3.0.0b2', ]
Pretty please? Our researchers want to use it, but it's a major pain (translation: impossible) to get it going for them.