xla
xla copied to clipboard
[WIP] Significantly reduce Docker image size
This PR brings the size of the GPU image down to 6.24 GB
from 17.2 GB
- Use local
xla
directory instead of cloning from github. This allows building an image from a local working copy - Separate build image from release image so we don't release build artifacts
- Break build steps up to take better advantage of caching to make debugging less painful. In particular, build TF in its own step because it takes forever
- Remove unnecessary dependencies
- Use smaller
python:X.Y-slim
base images
TODO:
- [ ] Add
torchvision
- [ ] Add shallow clone of tests, example models, etc
- [ ] Add support for patches
- [ ] Update this branch and test again
Needs follow-up:
- Create new GCB config for this image
- Collect wheels and upload to GCS
- Start using in e2e tests