func.return %2 : tensor<1x128x128x8xf32>

  // CHECK-LABEL: testDilatedConvWithNonZeroBasePadding
  // CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>)
  // CHECK-NEXT: [[RESULT:%.*]] = "tf.Conv2D"([[INPUT]], [[FILTER]]) <{dilations = [1, 2, 2, 1], padding = "SAME", strides = [1, 1, 1, 1]}> : (tensor<1x128x128x3xf32>, tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32>

// CHECK:      %[[SHAPE:[0-9]*]] = "tf.Shape"(%[[ARG_0]])
// CHECK-NEXT: %[[LAUNCH_A:[0-9]*]] = "tf_device.launch"
// CHECK-SAME: device = "a"
// CHECK-NEXT:   %[[OP_A:[0-9]*]] = "tf.opA"(%[[SHAPE]])
// CHECK-NEXT:   tf_device.return %[[OP_A]]
// CHECK:      %[[LAUNCH_B:[0-9]*]] = "tf_device.launch"
// CHECK-SAME: device = "b"
// CHECK-NEXT:   %[[OP_B:[0-9]*]] = "tf.opB"(%[[SHAPE]], %[[LAUNCH_A]])
// CHECK-NEXT:   tf_device.return %[[OP_B]]

                verify(programHost).compileSecondStageOf(
                    same(program),
                    same(scriptHost),
                    argThat { this.sourceHash == sourceHash && this.templateId == scriptTemplateId },
                    same(ProgramKind.TopLevel),
                    same(programTarget),
                    same(accessorsClassPath)
                )

                verifyNoMoreInteractions()

		}

	case *Interface:
		// Two interface types are identical if they describe the same type sets.
		// With the existing implementation restriction, this simplifies to:
		//
		// Two interface types are identical if they have the same set of methods with
		// the same names and identical function types, and if any type restrictions
		// are the same. Lower-case method names from different packages are always

  // CHECK-DAG: %[[VAL2:.+]] = "tfl.conv_2d"(%arg0, %[[VAL0]], %[[VAL1]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}>
  // CHECK: return %[[VAL2]]
  %0 = "tfl.pseudo_const"() {value = dense<42> : tensor<16x1x1x8xi8>} : () -> tensor<16x1x1x8xf32>

		}

	case *Interface:
		// Two interface types are identical if they describe the same type sets.
		// With the existing implementation restriction, this simplifies to:
		//
		// Two interface types are identical if they have the same set of methods with
		// the same names and identical function types, and if any type restrictions
		// are the same. Lower-case method names from different packages are always

    // CHECK-SAME: (tensor<2xi32>) -> tensor<1xi32>
    // CHECK: %[[DIM0:.*]] = mhlo.reshape %[[SLICE0]] : (tensor<1xi32>) -> tensor<i32>

    // CHECK: %[[SLICE1:.*]] = "mhlo.slice"(%[[ARG2]])
    // CHECK-DAG-SAME: start_indices = dense<1> : tensor<1xi64>
    // CHECK-DAG-SAME: limit_indices = dense<2> : tensor<1xi64>
    // CHECK-DAG-SAME: strides = dense<1> : tensor<1xi64>

  // CHECK: [[rendezvous_key_base:%.*]], [[result_future:%.*]] = tf_mlrt_tpu.compile_and_execute
  // CHECK-SAME: constant_operand_indices = array<i32: 2>
  // CHECK-SAME: num_operands = 4
  // CHECK-SAME: operands_with_static_shape = array<i32: 0, 1, 3>

If you can make sure that the same build cache is used for all builds then the task will always have the same outputs for the same inputs by design of the build cache.
Going down this road can lead to different problems with cache misses for incremental builds as described above.

  %3 = "tf.AddV2"(%2, %1): (tensor<15x28x28x1xf32>, tensor<1xf32>) -> tensor<15x28x28x1xf32>
  func.return %2 : tensor<15x28x28x1xf32>