Update app.py

app.py

@@ -5,6 +5,38 @@ import time
 from ccd import ccd_eval, run_eval
 from libra.eval.run_libra import load_model
 
+# =========================================
+# Safe Libra Hook: fallback to CPU if no CUDA
+# =========================================
+import torch
+import libra.model.builder as builder
+
+_original_load_pretrained_model = builder.load_pretrained_model
+
+def safe_load_pretrained_model(model_path, model_base=None, model_name=None, **kwargs):
+    print("[INFO] Applying safe device fallback hook for Libra...")
+    tokenizer, model, image_processor, context_len = _original_load_pretrained_model(
+        model_path, model_base, model_name, **kwargs
+    )
+
+    if torch.cuda.is_available():
+        device, dtype = "cuda", torch.float16
+        print("[INFO] GPU detected: using float16 precision.")
+    else:
+        device, dtype = "cpu", torch.float32
+        print("[WARN] No GPU detected. Running Libra on CPU with float32.")
+
+    try:
+        vision_tower = model.get_vision_tower()
+        vision_tower.to(device=device, dtype=dtype)
+        print(f"[INFO] Vision tower moved to {device} ({dtype})")
+    except Exception as e:
+        print(f"[WARN] Could not move vision tower: {e}")
+
+    return tokenizer, model, image_processor, context_len
+
+builder.load_pretrained_model = safe_load_pretrained_model
+
 
 # =========================================
 # Global Configuration
@@ -201,6 +233,12 @@ def main():
         gr.Markdown("""
         # 📷 CCD: Mitigating Hallucinations in Radiology MLLMs via Clinical Contrastive Decoding
         ### [Project Page](https://x-izhang.github.io/CCD/) | [Paper](https://arxiv.org/abs/2509.23379) | [Code](https://github.com/X-iZhang/CCD) | [Models](https://huggingface.co/collections/X-iZhang/libra-6772bfccc6079298a0fa5f8d)
+        
+        **🚨 Performance Warning**
+        
+        The demo is currently running on **CPU**, and a single inference takes approximately **500 seconds**.
+        To achieve optimal performance and significantly reduce inference time, **GPU** is required for effective operation. 
+        For more details, please refer to the [launch demo locally](https://github.com/X-iZhang/CCD#gradio-web-interface).
         """)
 
         with gr.Tab("✨ CCD Demo"):