diff --git a/pkgs/development/python-modules/llama-cpp-python/default.nix b/pkgs/development/python-modules/llama-cpp-python/default.nix
index 08597d833e768..27781bd0c2c2e 100644
--- a/pkgs/development/python-modules/llama-cpp-python/default.nix
+++ b/pkgs/development/python-modules/llama-cpp-python/default.nix
@@ -9,6 +9,7 @@
   # nativeBuildInputs
   cmake,
   ninja,
+  autoAddDriverRunpath,
 
   # build-system
   pathspec,
@@ -73,13 +74,13 @@ buildPythonPackage rec {
     # -mcpu, breaking linux build as follows:
     #
     # cc1: error: unknown value ‘native+nodotprod+noi8mm+nosve’ for ‘-mcpu’
-    "-DGGML_NATIVE=off"
-    "-DGGML_BUILD_NUMBER=1"
+    (lib.cmakeBool "GGML_NATIVE" false)
+    (lib.cmakeFeature "GGML_BUILD_NUMBER" "1")
   ]
   ++ lib.optionals cudaSupport [
-    "-DGGML_CUDA=on"
-    "-DCUDAToolkit_ROOT=${lib.getDev cudaPackages.cuda_nvcc}"
-    "-DCMAKE_CUDA_COMPILER=${lib.getExe cudaPackages.cuda_nvcc}"
+    (lib.cmakeBool "GGML_CUDA" true)
+    (lib.cmakeFeature "CUDAToolkit_ROOT" "${lib.getDev cudaPackages.cuda_nvcc}")
+    (lib.cmakeFeature "CMAKE_CUDA_COMPILER" "${lib.getExe cudaPackages.cuda_nvcc}")
   ];
 
   enableParallelBuilding = true;
@@ -87,6 +88,9 @@ buildPythonPackage rec {
   nativeBuildInputs = [
     cmake
     ninja
+  ]
+  ++ lib.optionals cudaSupport [
+    autoAddDriverRunpath
   ];
 
   build-system = [
@@ -125,7 +129,13 @@ buildPythonPackage rec {
     "test_real_llama"
   ];
 
-  pythonImportsCheck = [ "llama_cpp" ];
+  pythonImportsCheck = lib.optionals (!cudaSupport) [
+    # `libllama.so` is loaded at import time, and failing when cudaSupport is enabled as the cuda
+    # driver is missing in the sandbox:
+    # RuntimeError: Failed to load shared library '/nix/store/...-python3.13-llama-cpp-python-0.3.16/lib/python3.13/site-packages/llama_cpp/lib/libllama.so':
+    # libcuda.so.1: cannot open shared object file: No such file or directory
+    "llama_cpp"
+  ];
 
   passthru = {
     updateScript = gitUpdater {