Allocate more shared memory to attention kernel (#1154)

csrc/cuda_utils_kernels.cu

@@ -0,0 +1,14 @@
+int get_device_attribute(
+    int attribute,
+    int device_id)
+{
+    int device, value;
+    if (device_id < 0) {
+        cudaGetDevice(&device);
+    }
+    else {
+        device = device_id;
+    }
+    cudaDeviceGetAttribute(&value, static_cast<cudaDeviceAttr>(attribute), device);
+    return value;
+}