// CHECK-DAG: %[[input_7:.*]] = "tfl.dequantize"({{.*}}) : (tensor<2x4x!quant.uniform<i8<-127:127>:f32, 0.010094007169167826>>) -> tensor<2x4xf32>
// CHECK-DAG: %[[input_8:.*]] = "tfl.dequantize"({{.*}}) : (tensor<2x4x!quant.uniform<i8<-127:127>:f32, 0.018637238525030179>>) -> tensor<2x4xf32>

  %0 = stablehlo.constant dense<0xFF800000> : tensor<f32>  // -inf
  %1 = stablehlo.transpose %arg0, dims = [0, 3, 1, 2] : (tensor<1x16x16x4xf32>) -> tensor<1x4x16x16xf32>
  %2 = "stablehlo.reduce_window"(%1, %0) ({
  ^bb0(%arg1: tensor<f32>, %arg2: tensor<f32>):
      %3 = stablehlo.maximum %arg1, %arg2 : tensor<f32>
      stablehlo.return %3 : tensor<f32>

  // CHECK-NEXT:   %[[ID:.*]] = "tf_device.launch"() <{device = "TPU_REPLICATED_HOST_0"}> ({
  // CHECK-NEXT:     %[[IDINSIDE:.*]] = "tf.Identity"(%[[REPVAR]]) {ici_weight_distribution_mlir_bridge_marker = true} : (tensor<f32>) -> tensor<f32>
  // CHECK-NEXT:     tf_device.return %[[IDINSIDE]] : tensor<f32>
  // CHECK-NEXT:   }) : () -> tensor<f32>
  // CHECK-NEXT:   "tf_device.cluster"() ({

  %5 = "tfl.add"(%4, %arg3) {tac.device = "CPU", tac.inference_type = "QUANTIZED_INT8", fused_activation_function = "NONE"} : (tensor<1x384x128x!quant.uniform<i8:f32, 0.3:3>>, tensor<128x!quant.uniform<i8:f32, 0.2:-4>>) -> tensor<1x384x128x!quant.uniform<i8:f32, 0.3:-3>>
  func.return %5 : tensor<1x384x128x!quant.uniform<i8:f32, 0.3:-3>>
}

  %min = arith.constant dense<0.0> : tensor<f32>
  %max = arith.constant dense<255.0> : tensor<f32>
  %mini = "tf.Identity"(%min) : (tensor<f32>) -> tensor<f32>
  %maxi = "tf.Identity"(%max) : (tensor<f32>) -> tensor<f32>
  %minc = "tf.Cast"(%mini) : (tensor<f32>) -> tensor<f32>
  %maxc = "tf.Cast"(%maxi) : (tensor<f32>) -> tensor<f32>

// CHECK: return %[[UNIFORM_QUANTIZE_1]] : tensor<1x3x!quant.uniform<i8:f32, {{.*}}>>

// CHECK-PER-TENSOR: func.func private @quantized_dot_general_fn(%[[ARG_1:.+]]: tensor<1x2x!quant.uniform<i8:f32, {{.*}}>>, %[[ARG_2:.+]]: tensor<2x3x!quant.uniform<i8<-127:127>:f32, {{.*}}>>) -> tensor<1x3x!quant.uniform<i8:f32, {{.*}}>> attributes {_from_xla_call_module}

// CHECK-SAME:  [[VAR:%.*]]: tensor<f32>, [[MG:%.*]]: tensor<f32>, [[MS:%.*]]: tensor<f32>, [[MOM:%.*]]: tensor<f32>, [[LR:%.*]]: tensor<f32>, [[RHO:%.*]]: tensor<f32>, [[MOMENTUM:%.*]]: tensor<f32>, [[EPSILON:%.*]]: tensor<f32>, [[GRAD:%.*]]: tensor<f32>

    %8 = "mhlo.compare"(%arg1, %arg1) {comparison_direction = #mhlo<comparison_direction NE>} : (tensor<f32>, tensor<f32>) -> tensor<i1>
    %9 = mhlo.or %7, %8 : tensor<i1>
    %10 = "mhlo.select"(%9, %arg1, %arg3) : (tensor<i1>, tensor<f32>, tensor<f32>) -> tensor<f32>
    %11 = "mhlo.compare"(%arg1, %arg3) {comparison_direction = #mhlo<comparison_direction EQ>} : (tensor<f32>, tensor<f32>) -> tensor<i1>

  %min = arith.constant dense<0.0> : tensor<f32>
  %max = arith.constant dense<15.0> : tensor<f32>
  %mini = "tf.Identity"(%min) : (tensor<f32>) -> tensor<f32>
  %maxi = "tf.Identity"(%max) : (tensor<f32>) -> tensor<f32>
  %minc = "tf.Cast"(%mini) : (tensor<f32>) -> tensor<f32>
  %maxc = "tf.Cast"(%maxi) : (tensor<f32>) -> tensor<f32>

// CHECK-NEXT: [[res:%.+]] = "quantfork.dcast"([[conv]]) : (tensor<*x!quant.uniform<i8:f32, 0.023529411764705882:-128>>) -> tensor<*xf32>