Pytorch and cuda getting automatically upgraded during docker build.

[Anji-Builds]

I have noticed that the Pytorch and cuda version get upgraded to 2.3.1+cu121 as in the screenshot and when I tried to rebuild it is 2.4+cu121. I have used the exact Dockerfile only with minor changes, so I am unable to understand which line/what is causing the upgrade from 2.1.2+cu118. Also does it affect tiny-cuda-nn or any other working, if the finally docker container is a higher version than the recommended?

Could someone help me out to understand this?

[georghess]

Have you checked if the same thing happens when you run the provided Dockerfile?

[Anji-Builds]

@georghess it is the same Dockerfile. I had to add some additional package installations as otherwise I was getting errors. Also I couldnt do direct pip install of tiny-cuda-nn, therefore git cloned it and then installed it manually in the docker.

Also added --ignore-installed blinker while installing nerfstudio, due to error otherwise.

I have attached my Dockerfile for reference.
Dockerfile_neurad.txt

[georghess]

I just built an image from the dockerfile in the repo without any issues. Torch was not upgraded, tiny-cuda-nn built without problem.

We don't have the bandwidth to support any custom modifications of the provided dockerfile. However, I took a quick look at the file you sent and saw some differences like different torch version (2.1.2 instead of 2.0.1) some commented out install of an old tiny-cuda-nn version (v1.6 instead of the latest), later pinning of setuptools < 70.0. I'd suggest to start off with the given Dockerfile and only change what is truly needed for your additional packages, and try to install them as late as possible in the process.

[Anji-Builds]

I just built an image from the dockerfile in the repo without any issues. Torch was not upgraded, tiny-cuda-nn built without problem.

We don't have the bandwidth to support any custom modifications of the provided dockerfile. However, I took a quick look at the file you sent and saw some differences like different torch version (2.1.2 instead of 2.0.1) some commented out install of an old tiny-cuda-nn version (v1.6 instead of the latest), later pinning of setuptools < 70.0. I'd suggest to start off with the given Dockerfile and only change what is truly needed for your additional packages, and try to install them as late as possible in the process.

I had built it previously on Windows using docker desktop. However, when I built it on a Linux machine as per your instruction, the torch version doesn't change.

I wonder why there is a change while building on windows.