uv uses very agressive caching (https://docs.astral.sh/uv/concepts/cache/ https://docs.astral.sh/uv/concepts/cache/) but pytorch has many, many different combin

uv uses very agressive caching (https://docs.astral.sh/uv/concepts/cache/ https://docs.astral.sh/uv/concepts/cache/) but pytorch has many, many different combinations of packages for the same release. It's not uncommon to have to download pytorch many times, because some lib may specify one slightly different version than another: https://docs.astral.sh/uv/guides/integration/pytorch/#configuring-accelerators-with-environment-markers https://docs.astral.sh/uv/guides/integration/pytorch/#config... pytorch causes many headaches in the ecosystem, and a better handling of it is one thing their commercial offering, pyx, wants to fix: https://astral.sh/blog/introducing-pyx https://astral.sh/blog/introducing-pyx