Design a cache for large model weights on GPU nodes

Containers serving a model download its weights (around 40 GB) every time they start cold. Design a cache so that the second start is fast. A model version never changes once published, and it is stored as a manifest plus content-addressed chunks. Many GPU nodes may ask for the same model at the same moment.

Cover the layers of cache (regional and on each node), collapsing many requests for the same missing chunk into one download, checking integrity, evicting when disk runs low without pulling data out from under a running container, repairing corrupt chunks, and how you would tell whether the cache is healthy.