Fix the logic that causes an issue with philschmid/gemma-tokenizer-ch…

…atml tokenizer (#146)

The `setup_chat_format()` logic should not be applied to philschmid/gemma-tokenizer-chatml tokenizer, otherwise gemma models are trained w/o proper bos, eos tokens.

scripts/run_sft.py

@@ -122,7 +122,7 @@ def main():
 
     model = model_args.model_name_or_path
     # For ChatML we need to add special tokens and resize the embedding layer
-    if "<|im_start|>" in tokenizer.chat_template:
+    if "<|im_start|>" in tokenizer.chat_template and "gemma-tokenizer-chatml" not in tokenizer.name_or_path:
         model = AutoModelForCausalLM.from_pretrained(model_args.model_name_or_path, **model_kwargs)
         model, tokenizer = setup_chat_format(model, tokenizer)
         model_kwargs = None