wip

Signed-off-by: Raphael Glon <oOraph@users.noreply.github.com>

app.py

@@ -57,8 +57,7 @@ def _ensure_loaded():
     _device = next(_model.parameters()).device
 
 
-# _ensure_loaded()
-
+_ensure_loaded()
 LOG.info("DEVICE %s", _device)
 
 
@@ -80,7 +79,7 @@ def generate_stream(message: str, history: List[Tuple[str, str]]):
     """
 
     # FIXME: check the memory footprint doing so. We should rather do this before the spaces wrapper...
-    _ensure_loaded()
+    # _ensure_loaded()
 
     messages = _history_to_messages(history) + [{"role": "user", "content": message}]
     inputs = _tokenizer.apply_chat_template(