Update handler.py

handler.py

@@ -224,6 +224,35 @@ class EndpointHandler:
         )
         self.model.eval()
 
+         def _patch_forward(obj, label="model"):
+            try:
+                if not hasattr(obj, "forward"):
+                    return False
+                orig_forward = obj.forward
+
+                def patched_forward(*args, **kwargs):
+                    # Sessizce düşürülecek yeni anahtarlar
+                    kwargs.pop("cache_position", None)
+                    kwargs.pop("input_positions", None)
+                    return orig_forward(*args, **kwargs)
+
+                obj.forward = patched_forward
+                print(f"[hotfix] Patched forward on {label}")
+                return True
+            except Exception as e:
+                print(f"[warn] forward patch failed on {label}: {e}")
+                return False
+
+        # Ana modelde dene
+        _patch_forward(self.model, "self.model")
+
+        # Bazı sürümlerde forward zinciri iç modüle de gider
+        if hasattr(self.model, "model"):
+            _patch_forward(self.model.model, "self.model.model")
+        if hasattr(self.model, "base_model"):
+            _patch_forward(self.model.base_model, "self.model.base_model")
+        # =======================================================================
+
         # Model worker'dan: multimodal check
         self.is_multimodal = 'llava' in self.model_name.lower() or 'pulse' in self.model_name.lower()