diff --git a/vllm/attention/layer.py b/vllm/attention/layer.py
index 37f9a4b383ce9..a8e796a1eab63 100644
--- a/vllm/attention/layer.py
+++ b/vllm/attention/layer.py
@@ -310,7 +310,8 @@ class Attention(nn.Module, AttentionLayerBase):
             kv_sharing_target_layer_name,
             **extra_impl_args,
         )
-        self.backend = AttentionBackendEnum[self.attn_backend.get_name()]
+        backend_name = self.attn_backend.get_name()
+        self.backend = AttentionBackendEnum.__members__.get(backend_name)
         self.dtype = dtype
 
         # For cuda-alike (CUDA and ROCM) and cpu platforms, we control how