In this context, where it has been suggested that the relevant cached build products might be missing support for CUDA, I’d like to try re-evaluating the flake without using the nix-community substituters, to force local compilation, to see if a local build fares any better.
However, removing the substituters from the flake has no effect, as the required packages are already present in the local nix store, so the system does not need to download or build them.
How can I force the re-evaluation, and hence compilation of torch-bin in this context?
You mean rebuild/re-realise; it’s always re-evaluated. And if you’ve changed the inputs, it will get a separate store path, so realisation would happen again there as well. So, ultimately your issue is that the actual package expression used does not match the results you want; therefore, you’ll have to change the actual expression itself.
Keep in mind also, -bin packages generally refer to packages which are downloaded directly from the upstream, and at most patched/wrapped, but otherwise not built from source. So if the upstream binary is missing support for some feature, you’re basically SOL - you must use the non--bin variant.
Thank you for the terminology correction, I never remember these terms, and it bothers me.
My issue is that I want to try compiling it locally, rather than relying on what is found in the nix-community cache. Anyway, I hacked around it by switching to unstable nixpkgs. Unsurprisingly, the issue persists, which serves to discard that possibility.