> especially for GPU stuff you end up having to link against very specific details of the local architecture. Hrm. This doesn’t sound right to me. Any project

> especially for GPU stuff you end up having to link against very specific details of the local architecture. Hrm. This doesn’t sound right to me. Any project should target a particular version of Cuda and then the runtime machine simply needs to have that version available. Right? > a lot of people would like to have packages that use pre-installed dependencies that came with the system Those people are wrong. Everything these days requires Docker because it’s the only way to deploy software that can reliable not crash on startup. (This is almost entirely a Linux self induced problem)