Vincent Delbar
Vincent Delbar
@remicres I just sent you an email, there is something wrong with your r2.4 build on dockerhub, no way to import tf (core dump)...
I guess you should I you don't want to rebuild TF once again... Why did you build with branch r2.4 instead of default v2.4.1 ? It's probably an unstable branch...
Are you sure about 2.4.1 ? Because on Saturday you sent me this command by email : `docker build --network='host' -t otbtf:cpu --build-arg BASE_IMG=ubuntu:20.04 --build-arg TF=r2.4 --build-arg NUMPY_SPEC===1.17.4 .`
Indeed, for me this command prints nothing ! May be you messed up your tags ?
Interesting, could you show me your `lscpu` output ? It could be related to bazel `--config=opt`, may be you should avoid this to make sure your build is portable.
Strange, there's SSE4, AVX2 support everywhere so it shouldn't be an issue. I tried with your latest push and it seems ok (our OVH server is also Intel Xeon), which...
Oh I see, may be it is related to AVX512 instruction, which my Intel Xeon does not have, same as your laptop.
There's a dedicated package for AVX512, so this would also match with my explantation, it seems this instruction isn't compatible with regular CPUs... https://pypi.org/project/intel-tensorflow-avx512/
Yes, you need to remove either `--config=opt` from BZL_CONFIGS (should be enough) or `-march=native` from CC_OPT_FLAGS in build-env-tf.txt
OR you could remove `-march=native` and add regular opt like `--copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2` which should work with any recent hardware... I do believe it's only a problem for...