Bug Fix: `off` is parsed as `False` in yaml file, changed to `disabled` to avoid this.

src/llamafactory/hparams/model_args.py

@@ -97,7 +97,7 @@ class ModelArguments:
         default=None,
         metadata={"help": "Which scaling strategy should be adopted for the RoPE embeddings."},
     )
-    flash_attn: Literal["off", "sdpa", "fa2", "auto"] = field(
+    flash_attn: Literal["disabled", "sdpa", "fa2", "auto"] = field(
         default="auto",
         metadata={"help": "Enable FlashAttention for faster training and inference."},
     )

src/llamafactory/hparams/parser.py

@@ -102,6 +102,10 @@ def _verify_model_args(model_args: "ModelArguments", finetuning_args: "Finetunin
         if model_args.adapter_name_or_path is not None and len(model_args.adapter_name_or_path) != 1:
             raise ValueError("Quantized model only accepts a single adapter. Merge them first.")
 
+    # In case that `flash_attn` is set to `off` in the yaml file, and parsed as `False` afterwards.
+    if model_args.flash_attn == False:
+        raise ValueError("flash_attn should be \"disabled\", \"sdpa\", \"fa2\" or \"auto\".")
+
 
 def _check_extra_dependencies(
     model_args: "ModelArguments",

src/llamafactory/model/model_utils/attention.py

@@ -32,7 +32,7 @@ def configure_attn_implementation(config: "PretrainedConfig", model_args: "Model
     if model_args.flash_attn == "auto":
         return
 
-    elif model_args.flash_attn == "off":
+    elif model_args.flash_attn == "disabled":
         requested_attn_implementation = "eager"
 
     elif model_args.flash_attn == "sdpa":