From c222838e2d3619be5aaacf450fa21f77ee31ad9b Mon Sep 17 00:00:00 2001
From: Pascal <admin@serveurperso.com>
Date: Wed, 20 May 2026 23:21:27 +0200
Subject: [PATCH] server: re-inject subcommand when router spawns children
 under unified binary

---
 app/llama.cpp                  | 9 +++++++++
 tools/server/server-models.cpp | 8 ++++++++
 2 files changed, 17 insertions(+)

diff --git a/app/llama.cpp b/app/llama.cpp
index 10f909e2f3b..66257c56992 100644
--- a/app/llama.cpp
+++ b/app/llama.cpp
@@ -1,4 +1,5 @@
 #include <cstdio>
+#include <cstdlib>
 #include <string>
 #include <vector>
 
@@ -58,6 +59,14 @@ int main(int argc, char ** argv) {
 
     for (const auto & cmd : cmds) {
         if (matches(arg, cmd)) {
+
+            // router spawns children through this same binary, it needs the
+            // subcommand to relaunch as 'llama serve' and not bare options
+#ifdef _WIN32
+            _putenv_s("LLAMA_APP_CMD", cmd.name);
+#else
+            setenv("LLAMA_APP_CMD", cmd.name, 1);
+#endif
             return cmd.func(argc - 1, argv + 1);
         }
     }
diff --git a/tools/server/server-models.cpp b/tools/server/server-models.cpp
index ccf42320f77..47b6c2a4ec0 100644
--- a/tools/server/server-models.cpp
+++ b/tools/server/server-models.cpp
@@ -14,6 +14,7 @@
 #include <mutex>
 #include <condition_variable>
 #include <cstring>
+#include <cstdlib>
 #include <atomic>
 #include <chrono>
 #include <queue>
@@ -159,6 +160,13 @@ void server_model_meta::update_args(common_preset_context & ctx_preset, std::str
     // TODO: maybe validate preset before rendering ?
     // render args
     args = preset.to_args(bin_path);
+
+    // unified binary dispatches by subcommand, re-inject it right after the
+    // binary path so the child starts as 'llama serve ...' not 'llama ...'
+    const char * app_cmd = std::getenv("LLAMA_APP_CMD");
+    if (app_cmd != nullptr && app_cmd[0] != '\0' && !bin_path.empty()) {
+        args.insert(args.begin() + 1, app_cmd);
+    }
 }
 
 void server_model_meta::update_caps() {