Update

app.py

@@ -93,16 +93,9 @@ def predict_ner(sentence: str, model, tokenizer, id2label: dict, device="cpu", m
     enc = {k: v.to(device) for k, v in enc.items()}
 
     with torch.no_grad():
-        try:
-            out = model(**enc)
-            logits = out.logits if hasattr(out, "logits") else out
-        except TypeError:
-            # fallback for your custom BertSeqTagger-like model
-            h = model.transformer(**enc).last_hidden_state
-            if hasattr(model, "dropout"):
-                h = model.dropout(h)
-            logits = model.classification_head(h)
-
+        out = model(**enc)
+        logits = out.logits if hasattr(out, "logits") else out
+    
     pred_ids = logits.argmax(dim=-1)[0].tolist()
 
     word_ids = enc["input_ids"].new_zeros(enc["input_ids"].shape[1]).tolist()