// loaded, we can only check for availability after all arguments have been
// parsed.
//
// Furthermore, as `INSTANTIATE_TEST_SUITE_P` needs to be at global level and we
// don't want to have a `std::vector<std::string>` at global level, we use a
// static pointer to such a vector: we construct it via `SchemeVector()` below
// and when tests are instantiated we process it using `GetSchemes()`.

  vector<AbstractTensorHandle*> theta_inputs(inputs.size());
  for (int i{}; i < inputs.size(); ++i) {
    theta_inputs[i] = inputs[i];
  }

  AbstractTensorHandle* theta =
      theta_inputs[input_index];  // parameter we are grad checking

  // Convert from AbstractTensor to TF_Tensor.
  TF_Tensor* theta_tensor;
  TF_RETURN_IF_ERROR(GetValue(theta, &theta_tensor));

  // Get number of elements and fill data.

  std::unique_ptr<char[]> buffer(new char[kPosixCopyFileBufferSize]);

  off_t offset = 0;
  int bytes_transferred = 0;
  int rc = 1;
  // When `sendfile` returns 0 we stop copying and let callers handle this.
  while (offset < size && rc > 0) {
    size_t chunk = size - offset;
    if (chunk > kPosixCopyFileBufferSize) chunk = kPosixCopyFileBufferSize;

      if (iter == input_nodes.end()) {
        // This node is not referenced in inputs. Add it to the body.
        body_nodes->push_back(node);
      } else {
        // This node is referenced in inputs. Currently, we place an
        // artificial restriction and require that when num_opers=-1, such
        // nodes must have a single output.
        if (node->num_outputs() != 1) {
          return InvalidArgument(

    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.

  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);
  }

  TFE_ContextSetServerDef(ctx, 0, serialized.data(), serialized.size(), status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Use large matrices so that RPCs don't return before we get a chance
  // to call TFE_DeleteContext.
  TFE_TensorHandle* h0_task0 = TestMatrixTensorHandle100x100(ctx);
  TFE_TensorHandle* h1_task0 = TestMatrixTensorHandle100x100(ctx);
  const char remote_device_name[] =

}

TEST(PARALLEL_DEVICE, TestCollectiveSync) { TestCollective(/*async=*/false); }

// Note that ops on the parallel device currently don't execute
// asynchronously. The test is just that we don't get deadlocks.
TEST(PARALLEL_DEVICE, TestCollectiveAsync) { TestCollective(/*async=*/true); }

void RegisterCollectiveMulFunction(TFE_Context* context,

    if (forward_input) {
      // Since the input was forwarded, we released the input handle right away
      // and hence expect the input to be forwarded to the return tensor.
      EXPECT_EQ(orig_ptr, TF_TensorData(t));
    } else {
      // In async mode we expect forwarding to work without releasing the input
      // handle since by the time the kernel is executed we have released the
      // handle in the client code.

  // `gradient_checker` only works with model that returns only 1 tensor.
  // Although, `ops::SparseSoftmaxCrossEntropyWithLogits` returns 2 tensors, the
  // second tensor isn't needed for computing gradient so we could safely drop
  // it.
  outputs[0] = loss;
  backprop->Unref();
  return absl::OkStatus();
}

Status BiasAddModel(AbstractContext* ctx,
                    absl::Span<AbstractTensorHandle* const> inputs,