Problem

The flake.nix includes references to llama-cpp.cachix.org with a comment claiming it's "Populated by the CI in ggml-org/llama.cpp", but this is misleading and caused confusion during setup.

Personal Experience

I spent over an hour debugging why llama.cpp builds were taking so long, assuming the cache was functional based on the documentation. Only after extensive investigation did I realize the cache is not being populated.

Evidence the cache is non-functional:

1. No CI workflow exists: Searched .github/workflows/ - no workflows populate the cachix cache

2. Cache is accessible but empty for recent builds:

   $ nix store info --store https://llama-cpp.cachix.org
   Store URL: https://llama-cpp.cachix.org  # ← accessible
   
   $ nix build github:ggml-org/llama.cpp#cuda --dry-run
   these 5 derivations will be built:  # ← llama-cpp must be built locally
     /nix/store/2jqbcad4ig5indz3j88mjf05sjdhk9sc-cuda_nvcc-12.4.99.drv
     /nix/store/3q7gy8qnivrsbwkp0a7980475klv8hr0-cuda_cccl-12.4.99.drv
     /nix/store/mhk9108dvvlafr6p5xs55wa0k2ymanry-cuda_cudart-12.4.99.drv
     /nix/store/q2lqgcsnwsp765jqck7walgw47q5g4b0-libcublas-12.4.2.65.drv
     /nix/store/78lqkcz6n4fc3ck5mq7v2pry0qv81iys-llama-cpp-cuda-0.0.0.drv  # ← NOT cached
   these 155 paths will be fetched (704.42 MiB download, 1403.71 MiB unpacked)

   Note: Dependencies are cached (155 paths, 704MB), but llama.cpp itself is not.

Solution

Remove the non-functional cache references entirely, leaving only the working cuda-maintainers.cachix.org cache that actually provides CUDA dependencies.

This prevents other users from going through the same debugging process I did. The cache can always be re-added later if CI gets set up to populate it.

Testing

• Verified no GitHub Actions workflows populate the cache
• Confirmed llama-cpp builds not available in cache (must be built locally)
• Tested cache accessibility (works, but empty for llama.cpp binaries)
• Dependencies are properly cached via cuda-maintainers cache
• Documentation-only change, no functional impact