Clamp inference sequence length to BERT limit

src/inference.py

@@ -186,11 +186,20 @@ class AspectPredictor:
         unk_id = int(getattr(self.tokenizer, "unk_token_id", 0) or 0)
         return torch.where(input_ids >= vocab_size, torch.full_like(input_ids, unk_id), input_ids)
 
+    def _max_inference_length(self) -> int:
+        bert = getattr(self.model, "bert", None)
+        config = getattr(bert, "config", None)
+        max_positions = int(getattr(config, "max_position_embeddings", cfg.MAX_LENGTH) or cfg.MAX_LENGTH)
+        tokenizer_max = int(getattr(self.tokenizer, "model_max_length", cfg.MAX_LENGTH) or cfg.MAX_LENGTH)
+        if tokenizer_max > 100000:
+            tokenizer_max = cfg.MAX_LENGTH
+        return max(8, min(int(cfg.MAX_LENGTH), max_positions, tokenizer_max))
+
     def predict(self, review_text: str, product_meta: Dict,
                 return_attention: bool = True) -> Dict:
         enc = self.tokenizer(
             review_text,
-            max_length=cfg.MAX_LENGTH,
+            max_length=self._max_inference_length(),
             truncation=True,
             padding="max_length",
             return_tensors="pt",
@@ -234,7 +243,7 @@ class AspectPredictor:
             chunk = reviews[i:i + batch_size]
             enc = self.tokenizer(
                 [r["review_text"] for r in chunk],
-                max_length=cfg.MAX_LENGTH,
+                max_length=self._max_inference_length(),
                 truncation=True,
                 padding="max_length",
                 return_tensors="pt",