%1 = "tf.PartitionedCall"(%0) {config = "", config_proto = "", executor_type = "", f = @composite_dequantize_uniform} : (tensor<200x100x300xi8>) -> tensor<200x100x300xf32>
    %2 = "tf.XlaGather"(%1, %arg0, %cst) {dimension_numbers = "\0A\02\00\01\12\01\00\1A\02\00\01 \01", indices_are_sorted = true} : (tensor<200x100x300xf32>, tensor<10x2xi32>, tensor<3xi64>) -> tensor<1x300x10xf32>
    return %2 : tensor<1x300x10xf32>
  }

def IsXlaGatherWithoutBatch :
  Constraint<CPred<"IsXlaGatherWithoutBatch($0, $1)">>;

// Create Slice op from XlaGather op without batch dimension.
def CreateSliceAndReshapeOpFromXlaGatherOpWithoutBatch : NativeCodeCall<
  "CreateSliceAndReshapeOpFromXlaGatherOpWithoutBatch($_builder, $_loc, $0...)">;

// Convert XlaGather op without batch to Slice op with above two functions.
def ConvertXlaGatherOpWithoutBatch : Pat<

// dimension_numbers: {
//   offset_dims: 0
//   collapsed_slice_dims: 1
//   start_index_map: 1
// }
func.func @xla_gather(%arg0: tensor<?x2xf32>, %arg1: tensor<1xi32>, %arg2: tensor<2xi32>) -> tensor<*xf32> {
  %0 = "tf.XlaGather"(%arg0, %arg1, %arg2) {device = "", dimension_numbers = "\0A\01\00\12\01\01\1A\01\01", indices_are_sorted = true} : (tensor<?x2xf32>, tensor<1xi32>, tensor<2xi32>) -> tensor<*xf32>
  func.return %0 : tensor<*xf32>
}

                          XlaCompileOnDemandOp);                               \
  REGISTER_KERNEL_BUILDER(                                                     \
      Name("XlaGather").HostMemory("slice_sizes").Device(DEVICE),              \
      XlaCompileOnDemandOp);                                                   \
  REGISTER_KERNEL_BUILDER(Name("XlaScatter").Device(DEVICE),                   \

    %cst_0 = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<200x100x300xf32>} : () -> tensor<200x100x300xf32>
    %cst = "tf.Const"() { value = dense<[1, 1, 300]> : tensor<3xi64> } : () -> tensor<3xi64>
    %0 = "tf.XlaGather"(%cst_0, %arg0, %cst) {dimension_numbers = "\0A\02\00\01\12\01\00\1A\02\00\01\20\01", indices_are_sorted = true} : (tensor<200x100x300xf32>, tensor<10x2xi32>, tensor<3xi64>) -> tensor<1x300x10xf32>

def : Pat<(TF_XlaReplicaIdOp),
          (TF_CastOp (MHLO_ReplicaIdOp), /*truncate=*/ConstBoolAttrFalse)>;

//===----------------------------------------------------------------------===//
// XlaGather op.
//===----------------------------------------------------------------------===//

def ToGatherDimNumsAttr : NativeCodeCall<"GetGatherDimNumsAttr($0, &$_builder)">;

                                         "XlaDynamicSlice",
                                         "XlaDynamicUpdateSlice",
                                         "XlaEinsum",
                                         "XlaGather",
                                         "XlaIf",
                                         "XlaKeyValueSort",
                                         "XlaPad",
                                         "XlaRecv",

  %0 = "tf.Const"() {value = dense<0> : tensor<1xi32>} : () -> tensor<1xi32>
  %1 = "tf.Const"() {value = dense<[1, 9, 23, 768]> : tensor<4xi32>} : () -> tensor<4xi32>
  %2 = "tf.XlaGather"(%arg0, %0, %1) {device = "", dimension_numbers = "\0A\04\00\01\02\03\1A\01\02", indices_are_sorted = false} : (tensor<1x9x104x768xf32>, tensor<1xi32>, tensor<4xi32>) -> tensor<?x?x?x?xf32>
  func.return %2 : tensor<?x?x?x?xf32>

  func.func @xla_gather(%arg0: tensor<1x1x9xi32>, %arg1: tensor<1xi32>) -> tensor<*xi32> {
    %cst = "tf.Const"() {value = dense<[1, 1, 8]> : tensor<3xi32>} : () -> tensor<3xi32>
    %0 = "tf.XlaGather"(%arg0, %arg1, %cst) {dimension_numbers = "\0A\03\00\01\02\1A\01\02", indices_are_sorted = true} : (tensor<1x1x9xi32>, tensor<1xi32>, tensor<3xi32>) -> tensor<*xi32>
    func.return %0 : tensor<*xi32>
  }

      "XlaCustomCall",
      "XlaCustomCallV2",
      "XlaDequantize",
      "XlaDot",
      "XlaDotV2",
      "XlaDynamicSlice",
      "XlaDynamicUpdateSlice",
      "XlaEinsum",
      "XlaGather",
      "XlaIf",
      "XlaKeyValueSort",
      "XlaOptimizationBarrier",
      "XlaPad",
      "XlaRecv",
      "XlaReduce",
      "XlaReducePrecision",
      "XlaReduceScatter",