Merge pull request #313 from OpenAccess-AI-Collective/tokenizer-llama2-embeddings

don't resize embeddings to multiples of 32x by default

README.md

@@ -322,6 +322,9 @@ tokenizer_type: AutoTokenizer
 trust_remote_code:
 # use_fast option for tokenizer loading from_pretrained, default to True
 tokenizer_use_fast:
+# resize the model embeddings when new tokens are added to multiples of 32
+# this is reported to improve training speed on some models
+resize_token_embeddings_to_32x:
 
 # whether you are training a 4-bit GPTQ quantized model
 gptq: true

src/axolotl/utils/models.py

@@ -301,7 +301,11 @@ def load_model(
             **model_kwargs,
         )
 
-    embeddings_len = math.ceil(len(tokenizer) / 32) * 32
+    embeddings_len = (
+        math.ceil(len(tokenizer) / 32) * 32
+        if cfg.resize_token_embeddings_to_32x
+        else len(tokenizer)
+    )
     model.resize_token_embeddings(embeddings_len)
 
     if (