Update app.py

app.py

@@ -40,6 +40,9 @@ def load_model():
     
     # Load base model - can't do this on the free tier - not enough memory
     # model = AutoModelForCausalLM.from_pretrained(base_model_name)
+
+    # ✅ Force CPU placement before moving anything to CUDA
+    torch.cuda.is_available = lambda: False  # 👈 Trick PyTorch to think CUDA isn't available at first
     
     model = AutoModelForCausalLM.from_pretrained(
         base_model_name,
@@ -50,8 +53,11 @@ def load_model():
     )
     
     # Load LoRA adapter
-    model = PeftModel.from_pretrained(model, lora_model_name)
+    model = PeftModel.from_pretrained(model, lora_model_name, device_map={"": "cpu"})
 
+    # ✅ Now, allow CUDA again and move everything to GPU
+    torch.cuda.is_available = lambda: True
+    
     # Move model to GPU *AFTER* loading LoRA to avoid CUDA init errors
     model = model.to("cuda" if torch.cuda.is_available() else "cpu")