Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Clients build failed when set --build-arg "PYVER=3.6"

See original GitHub issue

Will failed

sudo docker build -t tensorrtserver_clients --target trtserver_build --build-arg "PYVER=3.6" --build-arg "BUILD_CLIENTS_ONLY=1" .

but 3.5 works

sudo docker build -t tensorrtserver_clients --target trtserver_build --build-arg "PYVER=3.5" --build-arg "BUILD_CLIENTS_ONLY=1" .

Issue Analytics

State:
Created 5 years ago
Comments:12 (4 by maintainers)

Top GitHub Comments

2reactions

mrjackbocommented, Jan 2, 2019

I have also had problems building the TRTIS client wheel for targets other than ubuntu 16.04 with Python 3.5. With the old github repository this was easily possible, but now there are dependencies in the Dockerfile which make this more complicated.

The problem is in the build step using the nvidia/tensorflow image. This is an image based on ubuntu 16.04 with python 3.5
/opt/tensorflow/nvbuild.sh is executed in nvidia/tensorflow,
One can work around usage of nvbuild.sh and create a wheel for a different python version, e.g. by installing miniconda with some other python version inside nvidia/tensorflow. But I was not able to build a wheel usable under ubuntu 18.04. The problem is that libcurl changed from 16.04 to 18.04, so to make a version of the client wheel usable under 18.04, one would need a version of the nvidia tensorflow image, based on ubuntu 18.04

In my opinion (and this is a feature request), the build process for the clients should be separated from the build process of the server, and made much simpler. After all (correct me if I’m wrong), the clients do not depend on tensorflow, tensorflow-serving, pytorch, caffe2 etc. One should only need grpc, libcurl, and the protobuf service definitions and hence the clients should be compilable in an almost generic current linux image. The compilation should also be quite quick…

1reaction

deadeyegoodwincommented, Mar 7, 2019

With https://github.com/NVIDIA/tensorrt-inference-server/pull/132 we now have a separate Makefile.client (and Dockerfile.client for convenience) that is used to build the client libraries and examples. The main Dockerfile now builds just the server.

Give it a try with your target OS and Python version and send in pull requests to fix any issues you find.