diff --git a/src/h3_blackwell_runtime/vae_decoder.py b/src/h3_blackwell_runtime/vae_decoder.py
index bed1238..18ebb05 100644
--- a/src/h3_blackwell_runtime/vae_decoder.py
+++ b/src/h3_blackwell_runtime/vae_decoder.py
@@ -102,7 +102,12 @@ class Attention(nn.Module):
 
             output = optimized_attention(query, key, value, self.heads, skip_reshape=True)
         except Exception:
-            output = F.scaled_dot_product_attention(query, key, value).transpose(1, 2)
+            try:
+                from sageattention import sageattn
+
+                output = sageattn(query, key, value, is_causal=False, tensor_layout="HND", smooth_k=False).transpose(1, 2)
+            except Exception:
+                output = F.scaled_dot_product_attention(query, key, value).transpose(1, 2)
         return self.to_out(output.reshape(batch, sequence, -1).nan_to_num_(0.0))