Fix .args loading order for program mode detection

llama.cpp/main/main.cpp

@@ -199,16 +199,16 @@ int main(int argc, char ** argv) {
         __builtin_unreachable();
     }
 
-    enum Program prog = determine_program(argv);
-    if (prog == LLAMAFILER)
-        return lf::server::main(argc, argv);
-
     mallopt(M_GRANULARITY, 2 * 1024 * 1024);
     mallopt(M_MMAP_THRESHOLD, 16 * 1024 * 1024);
     mallopt(M_TRIM_THRESHOLD, 128 * 1024 * 1024);
     ShowCrashReports();
     argc = cosmo_args("/zip/.args", &argv);
 
+    enum Program prog = determine_program(argv);
+    if (prog == LLAMAFILER)
+        return lf::server::run(argc, argv, true);
+
     if (prog == SERVER)
         return server_cli(argc, argv);
 

llamafile/server/main.cpp

@@ -20,5 +20,5 @@
 int
 main(int argc, char* argv[])
 {
-    return lf::server::main(argc, argv);
+    return lf::server::run(argc, argv, false);
 }

llamafile/server/prog.cpp

@@ -35,7 +35,7 @@ namespace server {
 Server* g_server;
 
 int
-main(int argc, char* argv[])
+run(int argc, char* argv[], bool args_already_loaded)
 {
     llamafile_check_cpu();
     signal(SIGPIPE, SIG_IGN);
@@ -56,7 +56,8 @@ main(int argc, char* argv[])
     }
 
     // get config
-    argc = cosmo_args("/zip/.args", &argv);
+    if (!args_already_loaded)
+        argc = cosmo_args("/zip/.args", &argv);
     llamafile_get_flags(argc, argv);
 
     // initialize subsystems

llamafile/server/prog.h

@@ -4,7 +4,7 @@ namespace lf {
 namespace server {
 
 int
-main(int, char**);
+run(int, char**, bool);
 
 } // namespace server
 } // namespace lf