// CHECK: return %6 : tensor<1x5x5x3xf32>
}

// CHECK-LABEL: bias_adjust_duplicate_filter
func.func @bias_adjust_duplicate_filter(%arg0: tensor<1x5x5x2xf32>) -> (tensor<1x5x5x3xf32>, tensor<1x5x5x3xf32>) {
  %0 = "quantfork.stats"(%arg0) {
    layerStats = dense<[-1.28e-5, 1.27e-5]> : tensor<2xf32>
  } : (tensor<1x5x5x2xf32>) -> tensor<1x5x5x2xf32>

  %3 = "tfl.quantize"(%2) {qtype = tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>, volatile} : (tensor<1x1x1x3xf32>) -> tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>
  %4 = "tfl.dequantize"(%3) : (tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>) -> tensor<1x1x1x3xf32>
  %5 = "tfl.add"(%1, %4) {fused_activation_function = "NONE"} : (tensor<1x5x5x3xf32>, tensor<1x1x1x3xf32>) -> tensor<1x5x5x3xf32>

// CHECK: %[[TRANSPOSE_1:.+]] = stablehlo.transpose %[[BROADCAST_IN_DIM]], dims = [0, 2, 3, 1] : (tensor<1x4x5x5xf32>) -> tensor<1x5x5x4xf32>
// CHECK: %[[ADD:.+]] = stablehlo.add %[[CONV]], %[[TRANSPOSE_1]] : tensor<1x5x5x4xf32>
// CHECK: %[[MAX:.+]] = stablehlo.maximum %[[ADD]], %[[ZERO_CONST]] : tensor<1x5x5x4xf32>

  // CHECK: %[[v1:.*]] = "tf.Reshape"(%arg0, %[[cst_1]]) : (tensor<2x5x7xf32>, tensor<4xi64>) -> tensor<2x5x1x7xf32>
  // CHECK: %[[v2:.*]] = "tf.BatchMatMulV2"(%[[v1]], %[[v0]]) <{adj_x = false, adj_y = false}> : (tensor<2x5x1x7xf32>, tensor<2x5x7x3xf32>) -> tensor<2x5x1x3xf32>
  // CHECK: %[[v3:.*]] = "tf.Reshape"(%[[v2]], %[[cst_2]]) : (tensor<2x5x1x3xf32>, tensor<3xi64>) -> tensor<2x5x3xf32>

    }> {
      _collective_manager_ids = [], device = ""
    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    func.return %0: tensor<1x2x2x3xf32>
  }

  func.func private @some_func(%arg0: tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32> {
    return %arg0 : tensor<1x2x2x3xf32>
  }
}

// CHECK: module
// CHECK-NOT: tf.PartitionedCall
// CHECK-NOT: some_func

}

// -----

module {
  func.func @conv(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
    %weight = "tf.Const"() {value = dense<2.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
    %1 = "tf.PartitionedCall"(%arg0, %weight) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_conv2d_fn_1} : (tensor<1x2x2x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

    has_token_input_output = false, module = "", platforms = [],
    version = 5 : i64
  } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>
  return %0 : tensor<1x2x2x2xf32>
}

// CHECK-LABEL: func.func @qdq_for_conv_weight_empty
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x2x3xf32>) -> tensor<1x2x2x2xf32>

  %0 = "tf.Conv2D"(%arg0, %cst) {data_format = "NHWC", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x3xf32>) -> tensor<1x3x2x3xf32>
  %1 = "tf.AddV2"(%0, %cst_0) : (tensor<1x3x2x3xf32>, tensor<1x3x2x3xf32>) -> tensor<1x3x2x3xf32>
  func.return %1 : tensor<1x3x2x3xf32>
}

    %0 = "tf.GatherV2"(%arg0, %arg1, %arg2) {attr_map = "0:batch_dims", batch_dims = 0 : i64, device = ""} : (tensor<1024x3x4x3xf32>, tensor<1xi32>, tensor<i32>) -> tensor<1x3x4x3xf32>
    return %0 : tensor<1x3x4x3xf32>
  }

    return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {