Avoid activating GPU when we can't use the GPU

Loading the GPU module imposes about a second of latency. If the user
isn't passing the -ngl 35 flag and it isn't Apple Silicon, then don't
bother running all the GPU-related initialization routines.

llama.cpp/common.cpp

@@ -562,7 +562,7 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_gpu_layers = std::stoi(argv[i]);
-            if (params.n_gpu_layers == 0) {
+            if (params.n_gpu_layers <= 0) {
                 FLAG_gpu = LLAMAFILE_GPU_DISABLE;
             }
         } else if (arg == "--gpu-layers-draft" || arg == "-ngld" || arg == "--n-gpu-layers-draft") {

llama.cpp/main/main.cpp

@@ -124,6 +124,13 @@ int main(int argc, char ** argv) {
         __builtin_unreachable();
     }
 
+    if (!IsXnuSilicon() &&
+        (!has_argument(argc, argv, "-ngl") &&
+         !has_argument(argc, argv, "--gpu-layers") &&
+         !has_argument(argc, argv, "--n-gpu-layers"))) {
+        FLAG_gpu = LLAMAFILE_GPU_DISABLE;
+    }
+
     if (!has_argument(argc, argv, "--cli") &&
         (has_argument(argc, argv, "--server") ||
          (!has_argument(argc, argv, "-p") &&