Installing horovod from my project setup.py

[SimonKenoby]

Hello,

I'm currently working on a project using tensorflow and we recenlty decided to test horovod to overcome multi gpus training bottleneck. As the result were positive we decided to integrate horovod in our environement, which is deployed via setup.py.

We are using different machine, some with GPUs, some without GPUs and I cannot make it so horovod is installed with tensorflow support on machine that don't have GPUs.

Our setup.py looks like this

install_requires=[ "IPython", ... "nltk", "numba", "numpy==1.19.*", "sklearn", "tensorflow-addons==0.14.*", "tensorflow==2.4.*", "tox", "horovod", # keep this after tensorflow ],
Installing with
HOROVOD_WITH_TENSORFLOW=1 HOROVOD_WITH_GLOO=1 pip install . --no-cache
However, I run into erros, saying that either tensorflow library is not found or that horovod has not been build with tensorflow support.

Check-build return me this:
Available Frameworks:
[ ] TensorFlow
[X] PyTorch
[ ] MXNet

Available Controllers:
[ ] MPI
[ ] Gloo

Available Tensor Operations:
[ ] NCCL
[ ] DDL
[ ] CCL
[ ] MPI
[ ] Gloo

In addition, I have trying adding "horovod[tensorflow]" as install_requiers, and indeed it works, but then it install tensorflow 2.7 while what I want is tensorflow 2.4 (for some lib compatibility reason).

Do you have any suggestion?

Regards,
Simon.