namespace tensorflow {

class ImmediateExecutionContext;
class AbstractOpAttrs;

// Abstract interface to an operation.
class ImmediateExecutionOperation : public AbstractOperation {
 public:
  virtual void Clear() = 0;

  // Returns the inputs of this op.
  virtual absl::Span<ImmediateExecutionTensorHandle* const> GetInputs()
      const = 0;
  virtual Status SetInput(size_t index,

TF_CAPI_EXPORT extern void TFE_ContextSetServerDefWithTimeout(
    TFE_Context* ctx, int keep_alive_secs, const void* proto, size_t proto_len,
    int64_t init_timeout_in_ms, TF_Status* status,
    bool clear_existing_contexts);

// Set server def with retries and timeout. This is helpful for fault-tolerant
// initial connection in high-preemption environments, such as
// ParameterServerStrategy training.

      /*retries=*/0, status, /*clear_existing_contexts=*/false);
}

// Set server def with timeout.
TF_CAPI_EXPORT extern void TFE_ContextSetServerDefWithTimeout(
    TFE_Context* ctx, int keep_alive_secs, const void* proto, size_t proto_len,
    int64_t init_timeout_in_ms, TF_Status* status,
    bool clear_existing_contexts) {
  TFE_ContextSetServerDefWithTimeoutAndRetries(

  // TODO(tfrt-devs): Figure out a way to deprecate following features after
  // migrated to TFRT.
  //===--------------------------------------------------------------------===//
  // Clear pending nodes in thread executors and kernel caches.
  virtual void ClearCachesAndThreadExecutors() = 0;

  // Initialize the step resource container for a training step. This is used

  } else if (!absl::SimpleAtoi(content_length->second, &read)) {
    TF_SetStatus(status, TF_UNKNOWN, "Could not get content-length header");
    return -1;
  }
  // `TF_OUT_OF_RANGE` isn't considered as an error. So we clear it here.
  TF_SetStatus(status, TF_OK, "");
  TF_VLog(1, "Successful read of %s @ %u of size: %u", path.c_str(), offset,
          read);
  stream.read(buffer, read);
  read = stream.gcount();

                           const size_t* lengths, int num_values) override {
    if (strcmp(attr_name, tensorflow::kColocationAttrName) == 0) {
      op_->colocation_constraints.clear();
      for (int i = 0; i < num_values; ++i) {
        op_->colocation_constraints.emplace(static_cast<const char*>(values[i]),
                                            lengths[i]);
      }
    } else {

                                     serialized_server_def.size(),
                                     init_timeout_in_ms, status,
                                     /*clear_existing_contexts=*/false);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteContextOptions(opts);
  TF_DeleteStatus(status);
  return ctx;
}

                 const char* raw_device_name, TF_Status* status) {
  if (op_to_reset) {
    tensorflow::ImmediateExecutionOperation* op =
        tensorflow::unwrap(op_to_reset);
    op->Clear();
    status->status = op->Reset(op_or_function_name, raw_device_name);
  } else {
    TF_SetStatus(status, TF_INVALID_ARGUMENT,
                 "op_to_reset should not be nullptr");
  }
}

  // Make grad_func a gradient of func
  TF_GraphCopyFunction(host_graph_, func, grad_func, s_);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);

  // Verify that function and its grad are linked
  gdef.Clear();
  GetGraphDef(host_graph_, &gdef);
  std::vector<std::pair<string, string>> grads = GetGradDefs(gdef);
  ASSERT_EQ(1, grads.size());
  ASSERT_EQ("FooFunc", grads[0].first);
  ASSERT_EQ("MyGrad", grads[0].second);

    *   Add tf.keras.layers.AbstractRNNCell as the preferred implementation of
        RNN cell for TF v2. User can use it to implement RNN cell with custom
        behavior.
    *   Adding `clear_losses` API to be able to clear losses at the end of
        forward pass in a custom training loop in eager.
    *   Add support for passing list of lists to the `metrics` param in Keras
        `compile`.