fix: set model max length to either passed in or tokenizer value (#17)

Author: anhuong

tuning/sft_trainer.py

@@ -93,8 +93,10 @@ def train(
             "pad_token": "<pad>",
         })
 
-    model_max_length = tokenizer.model_max_length
+    model_max_length = min(train_args.model_max_length, tokenizer.model_max_length)
     logger.info(f"Model max length {model_max_length}")
+    if train_args.model_max_length > tokenizer.model_max_length:
+        logger.warning(f"model_max_length {model_max_length} exceeds tokenizer.model_max_length {tokenizer.model_max_length}, using tokenizer.model_max_length {tokenizer.model_max_length}")
 
     # TODO: we need to change this, perhaps follow what open instruct does?
     special_tokens_dict = dict()