return %2 : tensor<?x2xf32>
      }
      func.func private @composite_fn_1(%arg0: tensor<?x2xf32>, %arg1: tensor<2x2xf32>, %arg2: tensor<2xf32>) -> tensor<?x2xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {
        return %arg0 : tensor<?x2xf32>
      }
    }
  )mlir";

// CHECK-LABEL: func @testWhileCasts(%arg0: tensor<!tf_type.variant<tensor<1x3xf32>>>) -> tensor<!tf_type.variant<tensor<*xf32>>>
func.func @testWhileCasts(%arg0: tensor<!tf_type.variant<tensor<1x3xf32>>>) -> (tensor<!tf_type.variant<tensor<*xf32>>>) {
  %0 = "tf.While"(%arg0) {
    cond = @testWhileCond, body = @testWhileBody, is_stateless = false
  } : (tensor<!tf_type.variant<tensor<1x3xf32>>>) -> (tensor<!tf_type.variant<tensor<*xf32>>>)

}
func.func @_func(%arg0: tensor<2x4xf32>, %arg1: tensor<4x2xf32>) -> tensor<2x2xf32> {
  %0 = "tf.MatMul"(%arg0, %arg1) {_XlaSharding = "\08\03\1A\02\02\01\22\02\00\01"} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1 = "tf.Identity"(%0) : (tensor<2x2xf32>) -> tensor<2x2xf32>
  return %1 : tensor<2x2xf32>
}

// -----
// The following op sharding is used in the following test case:

        # CHECK: value = dense<{{\[\[}}1.000000e+00, 2.000000e+00, 3.000000e+00], [4.000000e+00, 5.000000e+00, 6.000000e+00]]> : tensor<2x3xf32>
      }
    }
  }
}
node {
  name: "y"
  op: "Const"
  device: "/device:CPU:0"
  attr {
    key: "dtype"
    value {
      type: DT_INT64
    }
  }
  attr {

  %outputs =  "tf.Const"() {device = "/job:localhost/replica:0/task:0/device:CPU:0", value = dense<0> : tensor<i32>} : () -> tensor<i32>...

// NOFALLBACK-LABEL: @xla_svd
func.func @xla_svd(%arg0: tensor<1x1xf32>) -> (tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>) {
  // NOFALLBACK: XlaSvd
  %s, %u, %v = "tf.XlaSvd"(%arg0) {max_iter = 1, epsilon = 1.0E-09 : f32, precision_config = ""} : (tensor<1x1xf32>) -> (tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>)
  func.return %s, %u, %v : tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>
}

  %1 = "tf.ReadVariableOp"(%0) {device = "/CPU:0"} : (tensor<!tf_type.resource<tensor<1x1xf32>>>) -> tensor<1x1xf32>
  %2:2 = "tf.Split"(%cst, %arg0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x4xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>)
  %3:2 = "tf.Split"(%cst, %2#0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x2xf32>) -> (tensor<1x1xf32>, tensor<1x1xf32>)

}

func.func @einsum_matmul(%arg0: tensor<7x9xf32>, %arg1: tensor<9x5xf32>) -> tensor<7x5xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "ae,ed->ad"}: (tensor<7x9xf32>, tensor<9x5xf32>) -> tensor<7x5xf32>
  func.return %0 : tensor<7x5xf32>
  // CHECK-LABEL: einsum_matmul
  // CHECK: %[[v0:.*]] = "tf.BatchMatMulV2"(%arg0, %arg1) <{adj_x = false, adj_y = false}> : (tensor<7x9xf32>, tensor<9x5xf32>) -> tensor<7x5xf32>

    %2:2 = "tf.RecvTPUEmbeddingActivations"() {_tpu_embedding_layer = "call1", config = "\0A\0B\0C\0D"} : () -> (tensor<2x2xf32>, tensor<4x4xf32>)
    "tf.SendTPUEmbeddingGradients"(%2#0, %2#1) {_tpu_embedding_layer = "call1", config = "\0A\0B\0C\0D", operandSegmentSizes = array<i32: 2, 0>} : (tensor<2x2xf32>, tensor<4x4xf32>) -> ()
    tf_device.return

  %23 = mhlo.add %21, %22 : tensor<4xf32>
  %24 = "mhlo.broadcast_in_dim"(%3) <{broadcast_dimensions = dense<> : tensor<0xi64>}> : (tensor<f32>) -> tensor<4xf32>
  %25 = mhlo.multiply %23, %24 : tensor<4xf32>
  %26 = "mhlo.broadcast_in_dim"(%2) <{broadcast_dimensions = dense<> : tensor<0xi64>}> : (tensor<f32>) -> tensor<4xf32>
  %27 = mhlo.divide %25, %26 : tensor<4xf32>
  %28 = mhlo.floor %27 : tensor<4xf32>