Drop attention_mask at inference: the checkpoint was trained without one, and generate()'s auto-mask degenerates the output

app.py

@@ -79,6 +79,20 @@ def _load():
                 model = model.to("cuda")
             model.eval()
             model.config.use_cache = False
+
+            # This checkpoint was instruction-tuned with attention_mask=None, so the model
+            # never saw a mask during training. generate() builds one automatically, which
+            # pushes it off-distribution and degenerates the output ("국가의 수도는 국가의
+            # 수도입니다..."). Dropping the mask restores the trained behaviour. Measured
+            # 2026-07-20: with mask -> degenerate, without -> correct on every probe.
+            _fwd = model.forward
+
+            def _forward_without_mask(*a, **kw):
+                kw.pop("attention_mask", None)
+                return _fwd(*a, **kw)
+
+            model.forward = _forward_without_mask
+
             _state.update(model=model, tok=tok, device=dev, status="ready", error=None)
         except Exception as e:  # honest degrade — landing page still works
             _state.update(status="error", error=str(e)[:500])