gpu_options->set_allow_growth(gpu_memory_allow_growth);

  (*config.mutable_device_count())["CPU"] = num_cpu_devices;

  // TODO(b/113217601): This is needed for EagerContext::runner_ to use a
  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the

      TF_SetStatusFromIOError(status, errno, dst);
      return;
    }
  } else if (S_ISDIR(st.st_mode)) {
    TF_SetStatus(status, TF_FAILED_PRECONDITION, "target path is a directory");
    return;
  }

  // We cannot rename directories yet, so prevent this.
  if (stat(src, &st) != 0) {
    TF_SetStatusFromIOError(status, errno, src);
    return;
  } else if (S_ISDIR(st.st_mode)) {

TEST(CAPI, TestRemoteExecuteUpdateServerDefResourceAccessAsync) {
  TestRemoteExecuteUpdateServerDefResourceAccess(true);
}

void TestRemoteExecuteUpdateServerDefWithFailures(bool async) {
  // Fail fast on GetStatus requests so we can get errors instead of timeout
  // when updating cluster with non-exsitent worker
  tensorflow::setenv("GRPC_FAIL_FAST", "TRUE", /*overwrite=*/1);

  tensorflow::ServerDef server_def = GetServerDef(2);

  // There are two ways to represent compact row-major data
  // 1) nullptr indicates tensor is compact and row-majored.
  // 2) fill in the strides array as the real case for compact row-major data.
  // Here we choose option 2, since some frameworks didn't handle the strides
  // argument properly.
  dlm_tensor->dl_tensor.strides = stride_arr->data();

  dlm_tensor->dl_tensor.byte_offset =

  // TODO(srbs): It seems like this is used only for performance optimization
  // and not for correctness. The only downside of keeping this 1 seems to be
  // that the gradient accumulation is unbounded and we will never
  // aggressively aggregate accumulated gradients to recover memory.
  // Revisit and fix.
  return 1;
}

// Consumes references to the tensors in the gradient_tensors list and returns

    }
    auto begin = data.begin();
    if (offset > pos) {
      // The block begins before the slice we're reading.
      begin += offset - pos;
    }
    auto end = data.end();
    if (pos + data.size() > offset + n) {
      // The block extends past the end of the slice we're reading.
      end -= (pos + data.size()) - (offset + n);
    }
    if (begin < end) {

    // Wait until all pending nodes have completed since they may have a
    // reference to default_cancellation_manager_. We ignore the status return
    // since we already have a bad status to propagate.
    TFE_ContextAsyncWait(context, await_status.get());
    // Reset the cancellation manager on a bad status. Otherwise we'll cancel
    // all future operations.
    default_cancellation_manager_ = absl::make_unique<CancellationManager>();
  }

    TF_DeleteOutputList(func_outputs);
  }

  /**
   * We traced so far this function:
   *
   *   def two_adds(a, b):
   *     my_add1 = a + b
   *     my_add2 = b + b
   *     return my_add1, my_add2
   *
   * Now we will execute this function with an eager context:
   *
   *   output1, output2 = two_adds(2.0, 3.0)
   *
   * and check that we got 5.0 and 6.0 as results.
   */

  // Build eager context.

    const TFE_OpAttrs* attributes, int expected_max_outputs,
    TF_Status* status) {
  absl::optional<std::vector<MaybeParallelTensorOwned>> result;
  // TODO(allenl): We should remove "TPU" from these op names at the very least,
  // or consider other ways of packing/unpacking parallel tensors.
  if (operation_name == std::string("TPUReplicatedInput")) {

  TF_EXPECT_OK(ReadCache(&cache4, want_filename, want_offset, want_n, &out));
  EXPECT_EQ(calls, 4);
}

TEST(RamFileBlockCacheTest, BlockAlignment) {
  // Initialize a 256-byte buffer.  This is the file underlying the reads we'll
  // do in this test.
  const size_t size = 256;
  std::vector<char> buf;
  buf.reserve(size);
  for (int i = 0; i < size; i++) {
    buf.push_back(i);
  }