Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    bhack
    @bhack
    @angerson Is your registry not trusted for the CI running in Google?
    Cause instead actions are working correctly tensorflow/addons#2515
    Austin Anderson
    @angerson
    You have a log from today? That one is from July 5th
    bhack
    @bhack
    I will try to request to Kokoro another run
    Austin Anderson
    @angerson
    bhack
    @bhack
    I have switched in tensorflow/addons#2515
    What's next in the roadmap
    Austin Anderson
    @angerson
    Next is getting the tf-nightly releases built with these containers, but I'm pretty backed up right now, sadly
    Austin Anderson
    @angerson
    Just a reminder, the next SIG Build community call is tomorrow, Tuesday, August 3, at 2pm Pacific. See here for the notes and agenda: http://bit.ly/tf-sig-build-notes
    There won't be a lot of news sadly, since all my time has been taken up by an internal Ubuntu 18 upgrade that took longer than I'd hoped :\
    bhack
    @bhack
    I saw that Keras RC2 depends on tf-nightly without a specific version/date
    Are we not going to be in sync between RC releases?
    Austin Anderson
    @angerson
    @mihaimaruseac ^
    bhack
    @bhack
    It seems that we need to manually force the Keras RC2
    We declare
    pipdeptree -p tensorflow-cpu | grep keras
    - keras [required: >=2.6.0rc1,<2.7.0,
    So every TF RC is compatible with every Keras RC/stable
    Probably we are going to restrict the required with the TF stable release
    Mihai Maruseac
    @mihaimaruseac
    sorry, saw this only now
    for someone that develops on keras, they need some TF installed. tf-nightly is fine.
    for someone that installs TF, keras comes as dependency.
    If you want to try it for TF just click the button on tensorflow/tensorflow#48679
    immagine.png
    bhack
    @bhack
    I've crated a new PR so you can better explore the Github codespaces with TF tensorflow/tensorflow#51439
    Austin Anderson
    @angerson
    The next SIG Build community call is today, Tuesday, September 7, at 2pm Pacific. See here for the notes and agenda: http://bit.ly/tf-sig-build-notes
    Subin Modeel
    @sub-mod
    @angerson can you send me the logs of CUDA 11 build failures on centos7
    Austin Anderson
    @angerson
    Which build failures are those? I don't remember mentioning that
    Do you mean the "CUDA 11.4 needs manylinux2014" errors?
    Here's the short part of that error:
    bazel-out/k8-opt/bin/_solib_local/_U@local_Uconfig_Ucuda_S_Scuda_Ccudart___Uexternal_Slocal_Uconfig_Ucuda_Scuda_Scuda_Slib/libcudart.so.11.0: undefined reference to `memcpy@GLIBC_2.14'
    bazel-out/k8-opt/bin/_solib_local/_U@local_Uconfig_Ucuda_S_Scuda_Ccudart___Uexternal_Slocal_Uconfig_Ucuda_Scuda_Scuda_Slib/libcudart.so.11.0: undefined reference to `clock_gettime@GLIBC_2.17'
    Austin Anderson
    @angerson
    I should be able to find longer logs if they'll be helpful for you
    Subin Modeel
    @sub-mod
    yes, I was talking about CUDA 11.4 needs manylinux2014" errors . longer logs would be helpful.
    Just to be clear these errors were seen on the ubuntu docker image right ? The one here https://github.com/tensorflow/build/tree/master/official_build_environments/master/linux
    Subin Modeel
    @sub-mod
    @mihaimaruseac @angerson why did we move from ubuntu18.04 to ubuntu20.04 in the build Dockerfile ?
    Of the bazel invocation, specifically
    Re: the Dockerfiles, jumping up to 20.04 was commonly requested and a quick way to drop Python 3.6
    And not directly -- we saw the errors on our VM builds, as we haven't adopted any Docker stuff yet, since I was off that project for many weeks trying to upgrade to Ubuntu 18 on our VMs (which turned out to have been a waste of time)
    But I think they would appear also on the Dockerfiles, if they were changed to use CUDA 11.4, but those are not currently
    bhack
    @bhack
    @angerson Have we checked this table?
    Austin Anderson
    @angerson
    We confirmed with Nvidia directly that CUDA 11.4 requires glibc newer than manylinux2010
    The one from manylinux2014 is new enough
    bhack
    @bhack
    I think It Is in that release note for Centos Is 7.x and glibc 2.17
    bhack
    @bhack
    @perfinion Can you take a look at the windows build log of tensorflow/tensorflow#52173 ?
    Austin Anderson
    @angerson
    The next SIG Build community call is tomorrow, Tuesday, October 5, at 2pm Pacific. See here for the notes and agenda: http://bit.ly/tf-sig-build-notes
    bhack
    @bhack
    @angerson I've just started a regular CPU build in docker run -it --rm quay.io/pypa/manylinux_2_24_x86_64 /bin/bash
    It is going ok I'am at [3,122 / 8,687]
    I don't know if you are interested or not cause I'have only a few core free on the build machine..
    bhack
    @bhack
    As you have more HW resources please test tensorflow/build#41 if it is in your roadmap. If not close it.