diff --git a/qwen_tts/core/models/modeling_qwen3_tts.py b/qwen_tts/core/models/modeling_qwen3_tts.py
index 3ce406d..a3eb30a 100644
--- a/qwen_tts/core/models/modeling_qwen3_tts.py
+++ b/qwen_tts/core/models/modeling_qwen3_tts.py
@@ -1868,6 +1868,11 @@ class Qwen3TTSForConditionalGeneration(Qwen3TTSPreTrainedModel, GenerationMixin)
         weights_only=True,
         **kwargs,
     ):
+        # Hotfix to enable passing the correct attn implementation which is stored in the config but not in kwargs
+        requested_attn_implementation = kwargs.pop("attn_implementation", None)
+        if requested_attn_implementation is None and config and config._attn_implementation:
+            requested_attn_implementation = config._attn_implementation
+
         model = super().from_pretrained(
             pretrained_model_name_or_path,
             *model_args,
@@ -1880,6 +1885,7 @@ class Qwen3TTSForConditionalGeneration(Qwen3TTSPreTrainedModel, GenerationMixin)
             revision=revision,
             use_safetensors=use_safetensors,
             weights_only=weights_only,
+            attn_implementation=requested_attn_implementation,
             **kwargs,
         )
         if not local_files_only and not os.path.isdir(pretrained_model_name_or_path):