Update app.py

app.py

@@ -19,18 +19,18 @@ print("Loading tokenizer...")
 tokenizer = AutoTokenizer.from_pretrained(LORA_PATH, token=hf_token)
 
 print("Loading base model...")
+# CRITICAL FIX: Force the model into CPU memory during global startup
 base_model = AutoModelForCausalLM.from_pretrained(
     BASE_MODEL,
     torch_dtype=torch.bfloat16,
+    device_map="cpu",  
     token=hf_token
 )
 
 print("Loading LoRA adapter...")
 model = PeftModel.from_pretrained(base_model, LORA_PATH, token=hf_token)
 
-# Best practice for ZeroGPU: Keep model in CPU RAM globally.
-# We will move it to the GPU inside the annotated function.
-print("Model loaded into memory!")
+print("Model loaded into CPU memory safely!")
 
 # ─── Grapheme Tokenizer & Constants ───────────────────────────────────────────
 INSTRUCTION = (