Test out ggml-org/llama.cpp#4605 with my local config

Author: philiptaron

flake.lock

@@ -39,6 +39,7 @@
       overlays = builtins.attrValues overlays;
       config.allowUnfree = true;
       config.hostPlatform = system;
+      config.cudaSupport = true;
     };
     x86_64-linux = import nixpkgs (mkConfig "x86_64-linux");
     aarch64-darwin = import nixpkgs (mkConfig "aarch64-darwin");

flake.nix

@@ -38,6 +38,9 @@ in
     });
   });
 
+  # On zebul, we use CUDA 12.3
+  cudaPackages = final.cudaPackages_12_3;
+
   # Work in progress: build wpa_supplicant from source
   #wpa_supplicant = prev.wpa_supplicant.overrideAttrs (prevAttrs: {
   #  src = prev.fetchgit {

overlays.nix

@@ -160,9 +160,7 @@
     nurl
 
     # `llama-cpp` is a set of programs for running LLMs locally
-    (llama-cpp-cuda.override {
-      cudaPackages = pkgs.cudaPackages_12_3;
-    })
+    llama-cpp
   ];
 
   users.users.philip.packages = with pkgs; [