// TFL lstm only supports time-majored inputs, so if it's not time-majored,
  // we will transpose the inputs and outputs.
  auto time_major_attr = func_op->getAttrOfType<BoolAttr>("tf.time_major");
  if (time_major_attr == nullptr) return failure();

  bool time_majored = time_major_attr.getValue();
  auto input_type = mlir::dyn_cast_or_null<RankedTensorType>(input.getType());
  if (!input_type) {

    }
  }

  if (!inputs.empty() || !outputs.empty()) {
    arg_nodes->resize(inputs.size());
    ret_nodes->resize(outputs.size());

    for (Node* n : GetOrderedNodes()) {
      // Handle inputs/arguments.
      auto input_it = inputs.find(n->name());
      if (input_it != inputs.end()) {
        (*arg_nodes)[input_it->second] = {n, 0};
      }

  kQuantizationOp,  // Quantization ops have input/output attr.
};

// For each op type, the following axis carries axis information:
// kDynamicRangeOp: rhs_quantization_axis will carry axis information.
// kUnaryOp: quantization_axis will carry axis information.
// kBinaryOp: Among {lhs, rhs, output}_quantization_axis, only check rhs.
// kQuantizationOp: Among {input, output}_quantization_axis, only check input.

  %10 = "tf.opK"() : () -> tensor<*xi16>
  %11 = "tf.opL"() : () -> tensor<*xi64>
  tf_device.replicate([%0, %1, %2] as %input0: tensor<*xi1>, %9 as %input1: tensor<*xi8>, %10 as %input2: tensor<*xi16>, [%3, %4, %5] as %input3: tensor<*xi32>, [%6, %7, %8] as %input4: tensor<*xf32>, %11 as %input5: tensor<*xi64>) {n = 3 : i32} {
    tf_device.return
  }
  func.return

// CHECK:      %[[OP_A:[a-z0-9]*]] = "tf.opA"

// CHECK-SAME: %[[input_9]], %[[input_10]], %[[input_11]], %[[input_12]], %[[input_13]], %[[input_14]], %[[input_15]], %[[input_16]], %[[input_17]], %[[input_18]], %[[input_19]],
// CHECK-SAME: %[[input_20]], %[[input_21]], %[[input_22]], %[[input_23]])

is used instead.

Operands are replicated inputs and packed inputs.

replicated_inputs: each group of `n` inputs corresponds to an input for a single
individual replica and is mapped to a single region argument. Inside one group
the operands are matching in order the `devices` attribute. Each replicated
input must have compatible shapes and types.
packed_inputs: each input corresponds to an input broadcasted across all

// CHECK-SAME: %[[input_9]], %[[input_9]], %[[input_9]],
// CHECK-SAME: %[[input_10]], %[[input_11]], %[[input_12]], %[[input_13]],
// CHECK-SAME: %[[input_9]], %[[input_9]],
// CHECK-SAME: %[[input_14]], %[[input_15]],
// CHECK-SAME: %[[input_9]], %[[input_9]], %[[input_9]], %[[input_9]]) <{
// CHECK-SAME: asymmetric_quantize_inputs = false,
// CHECK-SAME: cell_clip = 1.000000e+01 : f32,

// RUN: tac-translate -input-mlir -output-mlir -device-specs=GPU %s -o - 2>&1 | FileCheck %s

module {
func.func @main(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> attributes {tf.entry_function = {inputs = "input0,input1,input2,input3", outputs = "output"}} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

        connection.queueIncoming(input1)
        connection.queueIncoming(input2)
        connection.queueIncoming(closeInput)
        received.await()
        daemonConnection.stop()

        then:
        1 * handler.onInput(input1)
        1 * handler.onInput(input2)
        1 * handler.onEndOfInput() >> { received.countDown() }
        0 * handler._
    }

# RUN: tf_tfl_translate -tf-input-arrays=input0,input1 \
# RUN:                  -tf-input-shapes=4:4 \
# RUN:                  -tf-input-data-types=DT_FLOAT,DT_FLOAT \
# RUN:                  -tf-output-arrays=Add \
# RUN:                  -tf-inference-type=DT_QUINT8 \
# RUN:                  -tf-input-min-values='-2,-3' \
# RUN:                  -tf-input-max-values='2,3' \
# RUN:                  --quant-stats=%s.stats \