support new special token #3420

Former-commit-id: f5c6a47f5193ab3a6c137580992bdcce0b31fdd5

src/llmtuner/model/adapter.py

@@ -157,6 +157,17 @@ def init_adapter(
             ):
                 raise ValueError("DoRA is not compatible with PTQ-quantized models.")
 
+            if model_args.resize_vocab and finetuning_args.additional_target is None:
+                input_embeddings = model.get_input_embeddings()
+                output_embeddings = model.get_output_embeddings()
+                module_names = set()
+                for name, module in model.named_modules():
+                    if module in [input_embeddings, output_embeddings]:
+                        module_names.add(name.split(".")[-1])
+
+                finetuning_args.additional_target = module_names
+                logger.warning("Vocab has been resized, add {} to trainable params.".format(",".join(module_names)))
+
             peft_kwargs = {
                 "r": finetuning_args.lora_rank,
                 "target_modules": target_modules,

src/llmtuner/model/loader.py

@@ -39,6 +39,8 @@ def _get_init_kwargs(model_args: "ModelArguments") -> Dict[str, Any]:
 def load_tokenizer(model_args: "ModelArguments") -> "PreTrainedTokenizer":
     r"""
     Loads pretrained tokenizer.
+
+    Note: including inplace operation of model_args.
     """
     init_kwargs = _get_init_kwargs(model_args)
     try:
@@ -57,6 +59,16 @@ def load_tokenizer(model_args: "ModelArguments") -> "PreTrainedTokenizer":
             **init_kwargs,
         )
 
+    if model_args.new_special_tokens is not None:
+        num_added_tokens = tokenizer.add_special_tokens(
+            dict(additional_special_tokens=model_args.new_special_tokens),
+            replace_additional_special_tokens=False,
+        )
+        logger.info("Add {} to special tokens.".format(",".join(model_args.new_special_tokens)))
+        if num_added_tokens > 0 and not model_args.resize_vocab:
+            model_args.resize_vocab = True
+            logger.warning("New tokens have been added, changed `resize_vocab` to True.")
+
     patch_tokenizer(tokenizer)
     return tokenizer
 

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.")