CUDA dependency make Docker Image bloated

deployment
1

According to Docker hub , the Docker image size increased from 2.76GB (1.12.1) to 5.29 GB (1.13.0).

I also checked the package size installed by pip, there seems not much change in wheel size (776.3MB to 890.1MB). However, we installed these oversized packages:

nvidia-cublas-cu11==11.10.3.66 (317.1 MB)
nvidia-cuda-nvrtc-cu11==11.7.99
nvidia-cuda-runtime-cu11==11.7.99 (557.1 MB)
nvidia-cudnn-cu11==8.5.0.96

So what do these do? And could we remove these dependencies?

2

These “oversized” packages were previously shipped in the PyTorch wheel directly, so the overall size shouldn’t change. Which exact container are you using which increased the size?

4

For all those packages, I install from PyPi directly

These are official Docker image from PyTorch size compare

image
image1178×71 10.8 KB

5

Thanks for the follow-up. Where did you find:

as these packages are using in the CUDA 11.7 wheels while your containers are using 11.3 and 11.6.
pip list | grep nvidia also doesn’t return anything in the 11.6 container, so are you mixing up a few containers?

6

Sorry for such a confusing message.

There are two separate problems actually.

  1. Docker image size increases a lot from PyTorch 1.12 to 1.13 (as shown in the above image).
  2. The python environment installed torch 1.13 is much larger than 1.12. I ran the pip install in a separate virtual environment to separately compare these two versions. To get the above package:
# Create virtual environment
python3.7 -m venv venv
source venv/bin/activate

# Should be install torch 1.13
pip install torch torchvision

# listing package
pip freeze | grep nvidia

I list these packages out because I think that they are the root problem that makes PyTorch become too big.