Wrap _prepare_4d_causal_attention_mask as a leaf function (#27236)

Wrap _prepare_4d_causal_attention_mask as a leaf function

src/transformers/models/llama/modeling_llama.py

@@ -40,6 +40,7 @@
     logging,
     replace_return_docstrings,
 )
+from ...utils.import_utils import is_torch_fx_available
 from .configuration_llama import LlamaConfig
 
 
@@ -48,6 +49,12 @@
     from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
 
 
+# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
+# It means that the function will not be traced through and simply appear as a node in the graph.
+if is_torch_fx_available():
+    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
+
+
 logger = logging.get_logger(__name__)
 
 _CONFIG_FOR_DOC = "LlamaConfig"