Horovod: Why does the Dockerfile uses 'devel' not the 'base'

[goswamig]

The Dockerfile is built on top of nvidia/cuda:9.0-devel-ubuntu16.04 while official Tensorflow has been built on top of nvidia/cuda:9.0-base-ubuntu16.04

Horovod Dockerfile: https://github.com/uber/horovod/blob/master/Dockerfile#L1
TensorFlow Dockerfile: https://github.com/tensorflow/tensorflow/blob/v1.12.0/tensorflow/tools/dockerfiles/dockerfiles/nvidia.Dockerfile#L38

[alsrgv]

@gautamkmr, to give us the flexibility to compile CUDA code if needed. We're planning to add MXNet to the image, and it needs to be compiled from source.

That said, if there's a desire for smaller docker images from the community, we could start producing a separate image for each framework, which would make TF and PyTorch images smaller.

[goswamig]

@alsrgv Thank you for prompt reply.

In fact I was trying to make it work on top of base or run time nvidia docker but it was not going well. I will give another shot. IMHO having smaller docker image will always be welcomed by community :)

At least TensorFlow it would make sense given that we are adding pip binary.

[goswamig]

Closing this for now, will revisit if I find any issue on building on top of base.