Know what you want - Dockerfile 7GB to 1GB

#docker #containers #python #devops

Encountered the classic big image issue when I built a docker image with multi staging.

Coming to what i was trying to achieve, I had this idea of hosting openai CLIP model for fun on gke so for that i was containerizing the following dependencies and app file, my requirement initially was

torch
torchvision
ftfy
regex
tqdm
fastapi
uvicorn
pillow
numpy<2
python-multipart

A lil bit of context here, the idea is to have the CLIP model that runs on CPU and doesn't need any CUDA/GPU support. So this requirement.txt installs the torch related packages with default GPU/CUDA which is not required and this made the image size to be 7gb

So like any other person I dug into it cos i was expecting the size to be around 2gb max. I found about the unwanted CUDA/GPU stuff being added while installing the torch packages.

So I updated the requirement.txt and mentioned that I just want CPU support and that just did it.

torch==2.2.0+cpu
torchvision==0.17.0+cpu
-f https://download.pytorch.org/whl/cpu/torch_stable.html
ftfy
regex
tqdm
fastapi
uvicorn
pillow
numpy<2
python-multipart

now the image size got drastically reduced to ~1GB from `~7GB

and if you want you can further drill it down by removing the tests and example directories from the site-package and that will probably save u 50mb ~ 60mb