diff --git a/CUDA/includes/attention.cuh b/CUDA/includes/attention.cuh
new file mode 100644
index 0000000..7feac08
--- /dev/null
+++ b/CUDA/includes/attention.cuh
@@ -0,0 +1,29 @@
+#pragma once
+
+#include "tensor.cuh"
+
+#include <cuda_runtime.h>
+
+namespace quadtrix {
+namespace cuda {
+
+Status attention_forward(
+    const TensorView& input_qkv,
+    TensorView preatt,
+    TensorView att,
+    TensorView output,
+    int num_heads,
+    cudaStream_t stream = nullptr);
+
+Status attention_backward(
+    const TensorView& grad_output,
+    const TensorView& input_qkv,
+    const TensorView& att,
+    TensorView grad_input_qkv,
+    TensorView grad_preatt,
+    TensorView grad_att,
+    int num_heads,
+    cudaStream_t stream = nullptr);
+
+}  // namespace cuda
+}  // namespace quadtrix