// that we'll execute at least two steps.
  builder.setInsertionPointAfter(non_tpu_caller_0);
  TF::StatefulPartitionedCallOp start_step_0;
  result = StartStep0(builder, loc, symbol_table, metadata_op, compilation_op,
                      C_0, orig_callers, loop_operands_0, start_step_0);
  if (failed(result)) return signalPassFailure();

  // Save the results of the forward_0 and core_tpu_0 calls by slicing them

    // CHECK: {{.*StatefulPartitionedCall.* f = @non_tpu.*}}
    // CHECK: {{.*StatefulPartitionedCall.* f = @start_step_0.*}}
    // CHECK: {{.*StatefulPartitionedCall.* f = @while_cond.*}}
    // CHECK: {{.*StatefulPartitionedCall.* f = @non_tpu.*}}
    // CHECK: {{.*StatefulPartitionedCall.* f = @start_step_1.*}}
    // CHECK: {{.*StatefulPartitionedCall.* f = @while_cond.*}}