fix tokenizer of chatglm2 (#2711)

Author: wangshuai09

fastchat/train/train.py

@@ -36,6 +36,9 @@
 @dataclass
 class ModelArguments:
     model_name_or_path: Optional[str] = field(default="facebook/opt-125m")
+    padding_side: str = field(
+        default="right", metadata={"help": "The padding side in tokenizer"}
+    )
 
 
 @dataclass
@@ -274,7 +277,7 @@ def train():
         model_args.model_name_or_path,
         cache_dir=training_args.cache_dir,
         model_max_length=training_args.model_max_length,
-        padding_side="right",
+        padding_side=model_args.padding_side,
         use_fast=False,
     )
     tokenizer.pad_token = tokenizer.unk_token