support new special token #3420

src/llmtuner/model/utils/embedding.py

@@ -42,9 +42,11 @@ def resize_embedding_layer(model: "PreTrainedModel", tokenizer: "PreTrainedToken
         current_embedding_size = model.get_input_embeddings().weight.size(0)
 
     if len(tokenizer) > current_embedding_size:
+        if getattr(model, "quantization_method", None):
+            raise ValueError("Cannot resize embedding layers of a quantized model.")
+
         if not isinstance(model.get_output_embeddings(), torch.nn.Linear):
-            logger.warning("Current model does not support resizing token embeddings.")
-            return
+            raise ValueError("Current model does not support resizing embedding layers.")
 
         model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=64)
         with context_maybe_zero3:

src/llmtuner/model/utils/rope.py

@@ -30,6 +30,10 @@ def configure_rope(config: "PretrainedConfig", model_args: "ModelArguments", is_
 
         current_max_length = getattr(config, "max_position_embeddings", None)
         if current_max_length and model_args.model_max_length > current_max_length:
+            logger.warning(
+                "Enlarge max model length from {} to {}.".format(current_max_length, model_args.model_max_length)
+            )
+            setattr(config, "max_position_embeddings", model_args.model_max_length)
             scaling_factor = float(math.ceil(model_args.model_max_length / current_max_length))
         else:
             logger.warning("Input length is smaller than max length. Consider increase input length.")