diff --git a/backend/app/services/stt.py b/backend/app/services/stt.py
index 37be580..952de86 100644
--- a/backend/app/services/stt.py
+++ b/backend/app/services/stt.py
@@ -13,6 +13,12 @@ from transformers import (
     AutoTokenizer,
     WhisperProcessor,
 )
+
+import transformers.models.llama.modeling_llama as _llama_mod
+
+if not hasattr(_llama_mod, "LLAMA_ATTENTION_CLASSES"):
+    _llama_mod.LLAMA_ATTENTION_CLASSES = _llama_mod.ALL_ATTENTION_FUNCTIONS
+
 from boson_multimodal.data_collator.higgs_audio_collator import HiggsAudioSampleCollator
 from boson_multimodal.data_types import ChatMLSample, AudioContent, Message
 from boson_multimodal.dataset.chatml_dataset import (