Update app.py

app.py

@@ -27,14 +27,14 @@ 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.
     """
-    # a) Load full model in 8‑bit
     model = AutoModelForImageTextToText.from_pretrained(
-        MODEL_ID,
-        trust_remote_code=True,
-        token=HF_TOKEN,
-        load_in_8bit=True,
-        device_map="auto"
-    )
+    MODEL_ID,
+    trust_remote_code=True,
+    token=HF_TOKEN,
+    torch_dtype=torch.float16,  # or torch.float32 if you prefer
+    device_map="auto"
+)
+
     device = next(model.parameters()).device
 
     def to_soap(text: str) -> str: