%cst_0 = "tf.Const"() {value = dense<2.000000e+00> : tensor<1024x2xf32>} : () -> tensor<1024x2xf32>
    %cst_1 = "tf.Const"() {value = dense<[2, 1024]> : tensor<2xi32>} : () -> tensor<2xi32>
    // Original identity preserved.
    %cst_identity = "tf.Identity"(%cst_0) {device = ""} : (tensor<1024x2xf32>) -> tensor<1024x2xf32>
    %0 = "tf.Reshape"(%cst_identity, %cst_1) : (tensor<1024x2xf32>, tensor<2xi32>) -> tensor<2x1024xf32>

tensor<32x1024xf32>, tensor<1024x1xf32>, tensor<1024xf32>, tensor<1024x1xf32>, tensor<i64>, tensor<i32>, tensor<32x1024xf32>, tensor<i64>, tensor<i64>, tensor<1024xf32>, tensor<32x1024xf32>, tensor<1024x1xf32>, tensor<1024xf32>, tensor<1024x1xf32>, tensor<i64>, tensor<i32>, tensor<32x1024xf32>, tensor<i64>, tensor<i64>, tensor<1024xf32>, tensor<32x1024xf32>, tensor<1024x1xf32>, tensor<1024xf32>, tensor<1024x1xf32>, tensor<i64>, tensor<i32>, tensor<32x1024xf32>, tensor<i64>, tensor<i64>, tensor<1024xf32>, tensor<32x1024xf32>,...

// CHECK: %[[v0:.*]] = "tf.Reshape"(%arg1, %[[cst]]) : (tensor<1024x8x128xf32>, tensor<2xi64>) -> tensor<1024x1024xf32>
// CHECK: %[[v1:.*]] = "tf.BatchMatMulV2"(%arg0, %0) <{adj_x = false, adj_y = false}> : (tensor<?x?x1024xf32>, tensor<1024x1024xf32>) -> tensor<?x?x1024xf32>
// CHECK: %[[v2:.*]] = "tf.Shape"(%arg0) : (tensor<?x?x1024xf32>) -> tensor<3xi32>

   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

    %1:4 = "tf.CustomAggregator"(%arg0) {calibration_method = 1 : i32, id = "0", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32} : (tensor<1024x1024xf32>) -> (tensor<1024x1024xf32>, tensor<f32>, tensor<f32>, tensor<*xi64>)

   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

}

// -----

module {
func.func private @func_20_GPU_FLOAT(%arg0: tensor<128x128xf32>, %arg1: tensor<3xi32>) -> tensor<1x128x128xf32> attributes {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_20"} {
  %0 = "tfl.reshape"(%arg0, %arg1) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<128x128xf32>, tensor<3xi32>) -> tensor<1x128x128xf32>
  func.return %0 : tensor<1x128x128xf32>
}

  %cst_weights = arith.constant dense<2.0> : tensor<1024x512xf32>
  %cst_bias = arith.constant dense<4.0> : tensor<1024xf32>

  %0 = "tfl.fully_connected" (%cst_input, %cst_weights, %cst_bias) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<512xf32>, tensor<1024x512xf32>, tensor<1024xf32>) -> tensor<1024xf32>

  func.return %0 : tensor<1024xf32>

// CHECK-NEXT:    %[[RESHAPED_BMM:.*]] = mhlo.reshape %[[BMM_0]]
// CHECK-NEXT:    return %[[RESHAPED_BMM]] : tensor<3x5x1x4xf32>
}


func.func @convert_dot_general_repeated(%arg0: tensor<1x1x1024xf32>, %arg1: tensor<1024x1024xf32>) -> tensor<1x1x1024xf32> {
  %0 = "mhlo.dot_general"(%arg0, %arg1) {
    dot_dimension_numbers = #mhlo.dot<
      lhs_batching_dimensions = [],
      lhs_contracting_dimensions = [2],
      rhs_batching_dimensions = [],

  // tf.XlaReduceScatter legalization
  //===--------------------------------------------------------------------===//
  // CHECK-LABEL: func @xla_reduce_scatter
  func.func @xla_reduce_scatter(%arg0: tensor<128x128xf32>) -> tensor<64x128xf32> {
      %cst = "tf.Const"() {value = dense<0> : tensor<i32>} : () -> tensor<i32>
      %cst_0 = "tf.Const"() {value = dense<[[0, 4], [1, 5], [2, 6], [3, 7]]> : tensor<4x2xi32>} : () -> tensor<4x2xi32>