update hparams

Former-commit-id: 1c4feac44192b1f540208837f5a530b0d3f5fb37

src/llamafactory/hparams/model_args.py

@@ -109,12 +109,6 @@ class ModelArguments:
         default=False,
         metadata={"help": "Enable shift short attention (S^2-Attn) proposed by LongLoRA."},
     )
-    efficient_packing: Optional[bool] = field(
-        default=None,
-        metadata={
-            "help": "Whether or not to pack the sequences without cross-contamination attention for efficient training."
-        },
-    )
     mixture_of_depths: Optional[Literal["convert", "load"]] = field(
         default=None,
         metadata={"help": "Convert the model to mixture-of-depths (MoD) or load the MoD model."},
@@ -232,6 +226,7 @@ class ModelArguments:
         self.compute_dtype: Optional["torch.dtype"] = None
         self.device_map: Optional[Union[str, Dict[str, Any]]] = None
         self.model_max_length: Optional[int] = None
+        self.block_diag_attn: bool = False
 
         if self.split_special_tokens and self.use_fast_tokenizer:
             raise ValueError("`split_special_tokens` is only supported for slow tokenizers.")
@@ -259,4 +254,5 @@ class ModelArguments:
         new_arg.compute_dtype = old_arg.compute_dtype
         new_arg.device_map = old_arg.device_map
         new_arg.model_max_length = old_arg.model_max_length
+        new_arg.block_diag_attn = old_arg.block_diag_attn
         return new_arg