Closed byronyi closed 3 years ago
Gentle ping @avijit-nervana
Ping @agramesh1 as well; looks like ldigo_p
got removed when upgrading MKLDNN to 0.18.
@byronyi There are two ways to use nGraph with TensorFlow - one is to compile nGraph with TensorFlow build - that you are attempting and the other is to build nGraph only and dynamically loading when running TensorFlow. The second approach is much easier to build and use. See the Build Option 2
here: https://github.com/tensorflow/ngraph-bridge#option-2-build-ngraph-bridge-with-binary-tensorflow-installation
Please try that and let us know if you run into any issues.
@byronyi There are two ways to use nGraph with TensorFlow - one is to compile nGraph with TensorFlow build - that you are attempting and the other is to build nGraph only and dynamically loading when running TensorFlow. The second approach is much easier to build and use. See the
Build Option 2
here: https://github.com/tensorflow/ngraph-bridge#option-2-build-ngraph-bridge-with-binary-tensorflow-installationPlease try that and let us know if you run into any issues.
Does that mean the Bazel option with the --opt=ngraph
no longer be supported? It might better be removed entirely if so.
We build nGraph with TF as we did customize our internal TF build. The binary approach isn't always the best choice as TF does not provide a stable ABI now (it's moving to that goal as in the modular TF RFC but it's gonna take some time to be implemented). See https://github.com/tensorflow/tensorflow/issues/27067 for a potential problem.
@byronyi
Is this issue present? If yes, please update the status with above suggestion. Is no, could you close it?
The nGraph build option seems to be removed and migrated to a separate package. Closing for now.
@byronyi
Thank you very much!
System information
Describe the problem
Any other info / logs
r1.14 branch: