- backward_api : abs_grad
  forward : tanh(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : abs_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           abs_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : acos_grad
  forward : acos(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : acos_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           acos_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : acosh_grad
  forward : acosh(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : acosh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           acosh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : add_grad
  forward : add(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  kernel :
    func : add_coo_coo_grad{sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo},
           add_csr_csr_grad{sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr}

- backward_api : addmm_grad
  forward : addmm(Tensor input, Tensor x, Tensor y, float alpha=1.0, float beta=1.0) -> Tensor(out)
  args : (Tensor input, Tensor x, Tensor y, Tensor out_grad, float alpha=1.0, float beta=1.0)
  output : Tensor(input_grad), Tensor(x_grad), Tensor(y_grad)
  kernel :
    func : addmm_csr_dense_grad {dense, sparse_csr, dense, dense -> dense, sparse_csr, dense},
           addmm_csr_csr_grad {sparse_csr, sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr, sparse_csr},
           addmm_coo_dense_grad {dense, sparse_coo, dense, dense -> dense, sparse_coo, dense},
           addmm_coo_coo_grad {sparse_coo, sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo, sparse_coo}

- backward_api : asin_grad
  forward : asin(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : asin_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           asin_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : asinh_grad
  forward : asinh(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : asinh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           asinh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : atan_grad
  forward : atan(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : atan_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           atan_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : atanh_grad
  forward : atanh(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : atanh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           atanh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : cast_grad
  forward : cast(Tensor x, DataType index_dtype, DataType value_dtype) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, DataType value_dtype)
  output : Tensor(x_grad)
  kernel :
    func : cast_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           cast_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
    data_type : out_grad

- backward_api : conv3d_coo_grad
  forward : conv3d_coo (Tensor x, Tensor kernel, int[] paddings, int[] dilations, int[] strides, int groups, bool subm, str key) -> Tensor(out), Tensor(rulebook), Tensor(counter)
  args : (Tensor x, Tensor kernel, Tensor out, Tensor rulebook, Tensor counter, Tensor out_grad, int[] paddings, int[] dilations, int[] strides, int groups, bool subm, str key)
  output : Tensor(x_grad), Tensor(kernel_grad)
  kernel :
    func : conv3d_coo_grad{sparse_coo, dense, sparse_coo, dense, dense, sparse_coo -> sparse_coo, dense}

- backward_api : divide_grad
  forward : divide(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  kernel :
    func : divide_coo_coo_grad{sparse_coo, sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo},
           divide_csr_csr_grad{sparse_csr, sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr}

- backward_api : divide_scalar_grad
  forward : divide_scalar (Tensor x, float scalar) -> Tensor(out)
  args : (Tensor out_grad, float scalar)
  output : Tensor(x_grad)
  invoke : divide_scalar(out_grad, scalar)

- backward_api : expm1_grad
  forward : expm1(Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : expm1_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           expm1_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : leaky_relu_grad
  forward : leaky_relu(Tensor x, float alpha) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float alpha)
  output : Tensor(x_grad)
  kernel :
    func : leaky_relu_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           leaky_relu_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : log1p_grad
  forward : log1p(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : log1p_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           log1p_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : masked_matmul_grad
  forward : masked_matmul(Tensor x, Tensor y, Tensor mask) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  kernel :
    func : masked_matmul_csr_grad{dense, dense, sparse_csr -> dense, dense}

- backward_api : matmul_grad
  forward : matmul(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  kernel :
    func : matmul_csr_dense_grad {sparse_csr, dense, dense -> sparse_csr, dense},
           matmul_csr_csr_grad {sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr},
           matmul_coo_dense_grad {sparse_coo, dense, dense -> sparse_coo, dense},
           matmul_coo_coo_grad {sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo}

- backward_api : maxpool_grad
  forward : maxpool(Tensor x, int[] kernel_sizes, int[] paddings, int[] dilations, int[] strides) -> Tensor(out), Tensor(rulebook), Tensor(counter)
  args : (Tensor x, Tensor rulebook, Tensor counter, Tensor out, Tensor out_grad, int[] kernel_sizes)
  output : Tensor(x_grad)
  kernel :
    func : maxpool_coo_grad {sparse_coo, dense, dense, sparse_coo, sparse_coo -> sparse_coo}

- backward_api : multiply_grad
  forward : multiply(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  kernel :
    func : multiply_coo_coo_grad{sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo},
           multiply_csr_csr_grad{sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr}

- backward_api : mv_grad
  forward : mv(Tensor x, Tensor vec) -> Tensor(out)
  args : (Tensor x, Tensor vec, Tensor out_grad)
  output : Tensor(x_grad), Tensor(vec_grad)
  kernel :
    func : mv_coo_grad{sparse_coo, dense, dense -> sparse_coo, dense},
           mv_csr_grad{sparse_csr, dense, dense -> sparse_csr, dense}

- backward_api : pow_grad
  forward : pow(Tensor x, float factor) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float factor)
  output : Tensor(x_grad)
  kernel :
    func : pow_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           pow_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : relu6_grad
  forward : relu6(Tensor x, float threshold) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  kernel :
    func : relu6_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           relu6_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : relu_grad
  forward : relu(Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : relu_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           relu_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : scale_grad
  forward : scale(Tensor x, float scale, float bias, bool bias_after_scale) -> Tensor(out)
  args : (Tensor out_grad, float scale)
  output : Tensor(x_grad)
  invoke : scale(out_grad, scale, 0.0, true)

- backward_api : sin_grad
  forward : sin(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : sin_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           sin_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : sinh_grad
  forward : sinh(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : sinh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           sinh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : softmax_grad
  forward : softmax(Tensor x, int axis=-1) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, int axis)
  output : Tensor(x_grad)
  kernel :
    func : softmax_csr_grad{sparse_csr, sparse_csr -> sparse_csr}

- backward_api : sparse_coo_tensor_grad
  forward : sparse_coo_tensor(Tensor values, Tensor indices, IntArray dense_shape) -> Tensor(out)
  args : (Tensor indices, Tensor out_grad)
  output : Tensor(values_grad)
  kernel :
    func : sparse_coo_tensor_grad{dense, sparse_coo -> dense}

- backward_api : sqrt_grad
  forward : sqrt(Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : sqrt_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           sqrt_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : square_grad
  forward : square(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : square_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           square_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : subtract_grad
  forward : subtract(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  kernel :
    func : subtract_coo_coo_grad{sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo},
           subtract_csr_csr_grad{sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr}

- backward_api : tan_grad
  forward : tan(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : tan_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           tan_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : tanh_grad
  forward : tanh(Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : tanh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
           tanh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}

- backward_api : to_dense_grad
  forward : to_dense(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : coo_to_dense_grad{sparse_coo, dense -> sparse_coo}

- backward_api : to_sparse_coo_grad
  forward : to_sparse_coo(Tensor x, int64_t sparse_dim) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : coo_to_dense { sparse_coo -> dense }

- backward_api : values_grad
  forward : values_coo(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  kernel :
    func : values_coo_grad{sparse_coo, dense-> sparse_coo}

- backward_api: fused_attention_grad
  forward : fused_attention_csr(Tensor query, Tensor key, Tensor value, Tensor sparse_mask, Tensor key_padding_mask, Tensor attn_mask) -> Tensor(out), Tensor(softmax)
  args: (Tensor query, Tensor key, Tensor value, Tensor softmax, Tensor out_grad)
  output : Tensor(query_grad), Tensor(key_grad), Tensor(value_grad)
  kernel :
    func : fused_attention_csr_grad{dense, dense, dense, sparse_csr, dense -> dense, dense, dense}
    layout : softmax
    data_type: query