✨[Feature] Provide wheel for Pytorch 1.11.0

[pvicente-sensei]

Is your feature request related to a problem? Please describe.

Installation with the lastest PyTorch release (Pytorch 1.11.0 + CUDA 11.3) requires master compilation.

Describe the solution you'd like

A new python wheel release for python to install Torch_TensorRT via pip3.

Describe alternatives you've considered

Using one of the docker containers instead of installing locally

[ncomly-nvidia]

Hi @pvicente-sensei thanks for the request! We are in the process of putting out a PyT 1.11 build in the coming weeks.

Re: alternative option - can you describe your preference for installing locally over using the docker containers on NGC?

[DuaneNielsen]

I'll give my 2c. Docker containers are great for deployment. They really suck for development.

For example, debugging code inside docker containers introduces a ton of new items to juggle that can go wrong. First I have to setup and get NVIDIA docker working. Then I have to setup the docker container volumes so I can have persistence, then I have to figure out how to get logging out and get the debugger hooked up. If I want to change something in the environment, I have to update and rebuild the container. If I want to plug a camera in... I have to figure out how to do that through docker.

That's all a lot more work than just making sure I have an NVIDIA driver and camera driver installed.

For large teams that can afford a docker guy, a CI-CD guy, and are managing a fleet of devices or production system, docker is great. But for a small team working on a prototype project, I'd prefer not...

[ncomly-nvidia]

Torch-TRT v1.1 is out, built on PyT 1.11. See the releases

[pvicente-sensei]

Hi @ncomly-nvidia
Sorry for the late reply and thanks for releasing v1.1.

I agree with @DuaneNielsen about being easier to prototype without docker.

[dignakov]

Hi @ncomly-nvidia I just wanted to comment on preference for pip vs docker. NGC docker containers can be hard to use with something like ROS, since it installs packages into system python or overlays if compiling from source - which don't play nicely with conda in ngc pytorch containers.

So for those use-cases, the containers are not really a good option. We can make a custom docker image starting with something like the Torch container, but would still need a convenient way to install torch-tensorrt. And for prototyping it's actually easier without docker.