Hi @dusty_nv, I used “jetson-containers build ollama”. The process created several images as follows:
ollama r36.3.0 b95a415240e9 3 hours ago 6.68GB
<none> <none> 430a1f9a2930 3 hours ago 9.33GB
ollama r36.3.0-cuda eaeb3b54e24d 3 hours ago 6.6GB
ollama r36.3.0-build-essential d26a5629e6ab 3 hours ago 723MB
The ollama:r36.3.0 image works, but it appears to depend on all those four images. At least if I try to delete them I get a message that they are being used (the ‘none’ does not seem to exist). Your images are about the size of my ollama:r36.6.0 and do not include the other ones, is there a way to remove those other images?
Hi @pontual, I believe that image currently showing <none> for you is a remnant of the multi-stage Dockerfile that ollama uses, wherein the binaries are copied into the final container. However that final container still depends on those previous two base images (like CUDA), so it won’t let you fully remove those (although you can untag them)
All docker layers are indexed by their SHA checksum and only stored once on disk, so those get shared between overlapping images with common bases, and is not using the total sum for storage. Sorry that it creates the extra entries in your docker images list (you should see mine), but that enables it to comprehensively test each stage during the build to ensure that GPU acceleration is properly working throughout the process (so that if/when something breaks it, you know where…)