func.func @lift_float_batch_matmul(%arg0: tensor<4x4x3xf32>) -> (tensor<4x4x3xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.0> : tensor<4x3x3xf32>} : () -> tensor<4x3x3xf32>
  %0 = "tf.BatchMatMulV2"(%arg0, %cst) {adj_x = false, adj_y = false, device = ""} : (tensor<4x4x3xf32>, tensor<4x3x3xf32>) -> tensor<4x4x3xf32>
  return %0 : tensor<4x4x3xf32>

  // CHECK-LABEL: test_batch_matmul_broadcast_to_einsum
  // CHECK: "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...kn->...mn"}> : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = false} : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  func.return %0: tensor<2x2x2xf32>
}

func.func @main(tensor<4x4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>) -> tensor<4x4x4xf32> {
// CHECK: {
// CHECK-NEXT:   version: 3,

// CHECK-NEXT: return %[[dq]] : tensor<2x2xf32>
}

// CHECK-LABEL: prepareStatistics
func.func @prepareStatistics(%arg0: tensor<8x4x3xf32>) -> tensor<8x4x3xf32> {
  %0 = "quantfork.stats"(%arg0) {
    layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>
  } : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32>
  %1 = "quantfork.stats"(%0) {
    layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>,
    axisStats = dense<[
      [-1.0, 1.0],

  "tf_device.cluster_func"(%1) {func = @cluster_func, use_spmd_for_xla_partitioning = true, num_cores_per_replica = 8 : i64} : (tensor<4x4x4xf32>) -> ()
  func.return
}

func.func @cluster_func(%arg0: tensor<4x4x4xf32>) {
  func.return
}

// -----

// Tests partitioned inputs/outputs with no sharding (via XLA SPMD) defaults to
// replicate sharding ("").

// CHECK-LABEL: func @partitioned_input_output

  %2 = "tf.Const"() {value = dense<0.0> : tensor<2x2xf32>} : () -> tensor<2x2xf32>
  %3 = "tf.Const"() {value = dense<0.0> : tensor<4x4xf32>} : () -> tensor<4x4xf32>
  "tf.SendTPUEmbeddingGradients"(%2, %3) {_tpu_embedding_layer = "call1", config = "\0A\0B\0C\0D", operandSegmentSizes = array<i32: 2, 0>} : (tensor<2x2xf32>, tensor<4x4xf32>) -> ()

  %3 = "tf.AddV2"(%arg0, %0): (tensor<4x4xf32>, tensor<1xf32>) -> tensor<4x4xf32>
  %4 = "tf.Log"(%3) {device = "/job:localhost/replica:0/task:0/device:GPU:0"}: (tensor<4x4xf32>) -> tensor<4x4xf32>

  // CHECK: %[[ADD1:.*]] = "tf.AddV2"
  // CHECK: %[[LOG1:.*]] = "tf.Log"(%[[ADD1]])
  %5 = "tf.AddV2"(%4, %1): (tensor<4x4xf32>, tensor<1xf32>) -> tensor<4x4xf32>

  %43 = "quantfork.stats"(%42) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x4x3xf32>) -> tensor<1x4x3xf32>
  %44 = "tfl.strided_slice"(%43, %1, %2, %3) {begin_mask = 7 : i32, ellipsis_mask = 0 : i32, end_mask = 5 : i32, new_axis_mask = 0 : i32, shrink_axis_mask = 0 : i32, offset = false} : (tensor<1x4x3xf32>, tensor<3xi32>, tensor<3xi32>, tensor<3xi32>) -> tensor<1x2x3xf32>

// CHECK: %[[CONV:.+]] = stablehlo.convolution(%[[TRANSPOSE_0]], %[[CONST]]) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = {{\[\[}}1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x4x4x8xf32>, tensor<3x3x8x8xf32>) -> tensor<1x4x4x8xf32>

  %3 = "tf.Squeeze"(%2) {squeeze_dims = [-3]} : (tensor<2x1x64x8xf32>) -> tensor<2x64x8xf32>
  %4 = "tf.BatchToSpaceND"(%3, %cst_1, %cst) : (tensor<2x64x8xf32>, tensor<1xi32>, tensor<1x2xi32>) -> tensor<1x128x8xf32>
  func.return %4 : tensor<1x128x8xf32>

  // CHECK-LABEL: testDilatedConv1DExpandH