feat-resize-tokenizer-add-new-tokens

unsloth_zoo/tokenizer_utils.py

@@ -73,15 +73,36 @@ def add_new_tokens(
     new_tokens = [],
     method = "mean",
     interpolation = 0.5,
+    resize_tokenizer: bool = True,
 ):
     """
     Smartly resizes the tokenizer and adds new tokens to the model.
     We also disregard untrained tokens by removing them from the mean calculation.
+
+    Parameters:
+    -----------
+    model : PreTrainedModel
+        The model to which new tokens will be added.
+    tokenizer : PreTrainedTokenizer
+        The tokenizer to be resized and updated with new tokens.
+    new_tokens : list or tuple, optional
+        A list or tuple of new tokens to be added. Default is an empty list.
+    method : str, optional
+        The method to use for adding new tokens. Can be "mean" or "interpolation". Default is "mean".
+    interpolation : float, optional
+        The interpolation factor to use when method is "interpolation". Should be between 0 and 1. Default is 0.5.
+    resized_tokenizer : bool, optional
+        A flag indicating whether the tokenizer should be resized. This is useful if the tokenizer is already 
+        resized with respect to the model. Default is True.
     """
     assert(isinstance(new_tokens, (list, tuple)))
-    assert(len(new_tokens) > 0)
     assert(method == "mean" or method == "interpolation")
     assert(interpolation >= 0 and interpolation <= 1)
+    # check if new tokens are not empty only if resized_tokenizer is True
+    if resize_tokenizer and len(new_tokens) == 0:
+        raise ValueError(
+            "Unsloth: You are trying to add new tokens to the model and tokenizer, but the new tokens list is empty."
+        )
 
     # Check if tokens already exist
     overlapping_tokens = set(new_tokens) & set(tokenizer.vocab.keys())
@@ -108,8 +129,19 @@ def add_new_tokens(
     old_config_size   = model.config.vocab_size
 
     # Add tokens!
-    old_length = len(tokenizer)
-    tokenizer.add_tokens(new_tokens)
+    if resize_tokenizer:
+        old_length = len(tokenizer)
+        tokenizer.add_tokens(new_tokens)
+    else:
+        # Old length means the model vocabulary itself if tokenizer is already resized
+        old_length = old_config_size
+
+        # Sort by id
+        new_vocab_tokenizer = sorted(tokenizer.get_vocab().items(),key=lambda x: x[1])
+
+        # Get the new tokens
+        new_tokens = [x[0] for x in new_vocab_tokenizer[old_length:]]
+
     # Also resizes lm_head as well!
     model.resize_token_embeddings(len(tokenizer))