Update app.py

app.py

@@ -23,18 +23,17 @@ tokenizer = AutoTokenizer.from_pretrained(
 )
 
 def generate_and_export():
-    """
-    On button click: lazily load the 8‑bit model, generate 100 doc→SOAP pairs,
-    split 70/30, run inference & eval, write files, and return download links.
-    """
-    model = AutoModelForImageTextToText.from_pretrained(
-    MODEL_ID,
-    trust_remote_code=True,
-    token=HF_TOKEN,
-    torch_dtype=torch.float16,  # or torch.float32 if you prefer
-    device_map="auto"
+    out = model.generate(
+    **inputs,
+    max_new_tokens=400,
+    do_sample=True,
+    top_p=0.95,
+    temperature=0.1,
+    pad_token_id=processor.tokenizer.eos_token_id,
+    use_cache=False         # ← disable HybridCache
 )
 
+
     device = next(model.parameters()).device
 
     def to_soap(text: str) -> str: