//
// This is intended to be used when a peer failure is detected.
TF_CAPI_EXPORT extern void TFE_AbortCollectiveOps(TFE_Context* ctx,
                                                  TF_Status* status);

// Checks the health of collective ops peers. Explicit health check is needed in
// multi worker collective ops to detect failures in the cluster.  If a peer is
// down, collective ops may hang.

  // in the cache, and our current block is not block size, this likely means
  // we have inconsistent state within the cache. Note: it's possible some
  // incomplete reads may still go undetected.
  if (block->data.size() < block_size_) {
    Key fmax = std::make_pair(key.first, std::numeric_limits<size_t>::max());
    auto fcmp = block_map_.upper_bound(fmax);

// things like the available devices, resource manager etc.
// TFE_Context must outlive all tensor handles created using it. In other
// words, TFE_DeleteContext() must be called after all tensor handles have
// been deleted (with TFE_DeleteTensorHandle).
//
// TODO(ashankar): Merge with TF_Session?
typedef struct TFE_Context TFE_Context;

TF_CAPI_EXPORT extern TFE_Context* TFE_NewContext(

namespace {

// Managing context for the DLManagedTensor, will manage the lifetime of
// DLManagedTensor. When calling DLManagedTensor::deleter, it will notify the
// original framework of destruction, and this context will be deleted also.
struct TfDlManagedTensorCtx {
  TensorReference reference;
  std::vector<int64_t> shape;
  std::vector<int64_t> strides;
  DLManagedTensor tensor;

    it may slow down the tests for not running in parallel and may cause the
    test to timeout but it could be useful when you need to execute a single
    test or more in general your filtered/selected tests have a very low
    execution time and the sharding
    [could create an overhead on the test execution](https://github.com/bazelbuild/bazel/issues/2113#issuecomment-264054799).

};

// If `persistent_tape` is true, op_tape is not changed and none of the
// backwards functions are deleted.
// If `persistent_tape` is false, op_tape is cleared and backwards functions
// not needed for gradient computation are deleted. Backwards functions that
// are needed, are copied and returned in BackpropInitialState.
template <typename BackwardFunction, typename TapeTensor>

  memcpy(ret, debug_str.c_str(), *len + 1);
  return ret;
}

// On success, returns a set of TF_Function instances from `text_proto` of
// GraphDef type. These functions must be deleted by calling TF_DeleteFunction.
//
// If `mutate_proto_func` is non-NULL, run it over each FunctionDef proto,
// before creating a TF_Function out of the possibly mutated proto.

                      TF_Status* status) {
  GeneralAssignment("AssignVariableOp", context, value, status);
}

// Passed to `TF_NewTensor` to indicate how an array of floats should be
// deleted.
static void FloatDeallocator(void* data, size_t, void* arg) {
  delete[] static_cast<float*>(data);
}

// Creates a TFE_TensorHandle with value `v`.
TensorHandlePtr FloatTensorHandle(float v, TF_Status* status) {

    std::unique_ptr<ParallelTensor> t, TF_Status* status) {
  // The resulting TensorHandle owns an opaque pointer to "device memory", which
  // for a ParallelDevice is really a ParallelTensor. When the TensorHandle is
  // deleted, it will call ParallelTensorDeallocator to free the struct.
  ParallelTensor* t_released = t.release();
  TFE_CustomDeviceTensorHandleMethods handle_methods;
  handle_methods.num_dims = &ParallelTensorNumDims;