Update app.py

app.py

@@ -62,15 +62,25 @@ def _lazy_model_tok():
     global _tokenizer, _model
     if _tokenizer is not None and _model is not None:
         return _model, _tokenizer
+
     _tokenizer = AutoTokenizer.from_pretrained(STORYBOARD_MODEL, trust_remote_code=True)
+
+    # Choose a dtype that works both locally and on ZeroGPU
+    use_cuda = torch.cuda.is_available()
+    preferred_dtype = torch.float16 if use_cuda else torch.float32  # torch.bfloat16 is also fine if supported
+
     _model = AutoModelForCausalLM.from_pretrained(
         STORYBOARD_MODEL,
         device_map="auto",
-        dtype="auto",
+        torch_dtype=preferred_dtype,   # ✅ FIXED: use torch_dtype
         trust_remote_code=True,
+        use_safetensors=True
     )
+
+    # Ensure pad token to avoid warnings
     if _tokenizer.pad_token_id is None and _tokenizer.eos_token_id is not None:
         _tokenizer.pad_token_id = _tokenizer.eos_token_id
+
     return _model, _tokenizer
 
 def _prompt_with_tags(user_prompt: str, n_shots: int, default_fps: int, default_len: int) -> str: