Upload model.py with huggingface_hub

model.py

@@ -18,6 +18,7 @@ _llm = None
 _using_mock = False
 _model_checked = False
 _gen_lock = threading.Lock()
+_load_lock = threading.Lock()
 
 IM_START = "<|im_start|>"
 IM_END = "<|im_end|>"
@@ -72,10 +73,20 @@ def _load_model():
     if _model_checked:
         return
 
-    _model_checked = True
+    # Double-checked lock: the startup warm-up thread and the first real request
+    # can both reach here at once; without this they would load the model twice.
+    with _load_lock:
+        if _model_checked:
+            return
+        _do_load_model()
+
+
+def _do_load_model():
+    global _llm, _using_mock, _model_checked
     gguf = _find_gguf()
     if gguf is None:
         _using_mock = True
+        _model_checked = True
         return
 
     try:
@@ -107,6 +118,7 @@ def _load_model():
         print(f"[model] Could not load GGUF ({gguf}): {exc}")
         _llm = None
         _using_mock = True
+    _model_checked = True
 
 
 def _build_prompt(system_prompt: str, user_prompt: str) -> str: