Update app.py

app.py

@@ -8,6 +8,10 @@ from transformers import AutoProcessor, AutoModelForVision2Seq
 from PIL import Image
 import gradio as gr
 
+# Force Hugging Face offline globally
+os.environ["HF_HUB_OFFLINE"] = "1"
+os.environ["HF_DATASETS_OFFLINE"] = "1"
+
 # Normalization boundary for ScreenVLM
 NORM_SIZE = 500
 MODEL_ID = "docling-project/ScreenVLM"
@@ -19,6 +23,7 @@ class ScreenVLMEngine:
         
         # --- THE CORRECT PROCESSOR OVERRIDE (MATCHED EXACTLY) ---
         if hasattr(self.processor, "image_processor"):
+            # Set dictionary structure correctly to prevent subscriptable errors
             self.processor.image_processor.max_image_size = {"longest_edge": 99999}
             if hasattr(self.processor.image_processor, "size"):
                 self.processor.image_processor.size = {"longest_edge": 512}
@@ -44,7 +49,7 @@ class ScreenVLMEngine:
         model = AutoModelForVision2Seq.from_pretrained(
             MODEL_ID,
             torch_dtype=torch.float32,     
-            attn_implementation="eager",   
+            attn_implementation="eager",   # Matched exactly to prevent the SDPA crash
             low_cpu_mem_usage=True          
         )
         
@@ -54,6 +59,7 @@ class ScreenVLMEngine:
         )
         
         # Free unquantized weights from RAM immediately
+        import gc
         del model
         gc.collect()
         print("Model is AWAKE and ready in RAM.")
@@ -83,6 +89,7 @@ class ScreenVLMEngine:
         return elements
 
     def analyze(self, image: Image.Image):
+        # Gradio passes image directly as a PIL Image object
         orig_width, orig_height = image.size
         
         # Keep our CPU-friendly manual resizing, but scale to 512 max edge
@@ -106,7 +113,7 @@ class ScreenVLMEngine:
             generated_ids = self.model.generate(
                 **inputs, 
                 max_new_tokens=1024,
-                use_cache=True,                 
+                use_cache=True,                 # Speeds up CPU generation by 10x-20x
                 eos_token_id=100257,            
                 pad_token_id=100257
             )
@@ -127,7 +134,6 @@ engine = ScreenVLMEngine()
 def gradio_interface_fn(image):
     try:
         elements = engine.analyze(image)
-        # FIX: Serialize to string to prevent Pydantic/Gradio 4 validation clashes
         return json.dumps({"status": "success", "elements": elements})
     except Exception as e:
         return json.dumps({"status": "error", "message": str(e)})
@@ -136,10 +142,10 @@ def gradio_interface_fn(image):
 demo = gr.Interface(
     fn=gradio_interface_fn,
     inputs=gr.Image(type="pil", label="Upload Screenshot"),
-    outputs="text",  # FIX: Set to text to bypass Schema checking
+    outputs="text",  
     title="ScreenVLM Production Parser"
 )
 
 if __name__ == "__main__":
-    # FIX: Explicitly bind host and port for Hugging Face Docker environments
-    demo.queue().launch(server_name="0.0.0.0", server_port=7860)
+    # Let Hugging Face manage host and port dynamically (safer for default Space configs)
+    demo.queue().launch()