dl4rce/flaiwheel

What's changed

Major: CPU-only PyTorch (image size ~8GB → ~1GB)

Root cause found: sentence-transformers depends on torch, which defaults to the full CUDA build — pulling in ~7GB of NVIDIA libraries (nvidia-cublas, nvidia-cudnn, nvidia-cuda-runtime, etc.). On a CPU-only server these are completely unused but:

• Caused "no space left on device" during layer export on 30GB disks
• Made layer export take 127s+ (compressing 8GB of .so files)
• Made the image too large for small VMs

Fix: The Dockerfile now installs torch + torchvision from the PyTorch CPU-only index before the main pip install. When sentence-transformers is installed afterward, it sees torch already satisfied and skips the CUDA variant.

| Metric | Before | After |
|--------|--------|-------|
| Image size | ~8 GB | ~1 GB |
| Layer export | 127s+ | ~10s |
| Disk needed | 20GB+ | 5GB |
| Build failures | Frequent on small VMs | None |

Also: more aggressive pre-build cleanup

build_image() now runs docker builder prune -af, docker image prune -af, and docker container prune -f before every build — ensuring maximum free space regardless of leftover state from prior builds.

How to upgrade

curl -sSL https://raw.githubusercontent.com/dl4rce/flaiwheel/v3.8.0/scripts/install.sh | bash