*    dA = U
     *    dbias = reduceSum(U, dims = channel_dim)
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    DCHECK(upstream_grad);

    // Recover data format from forward pass for gradient.
    std::string data_format;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("data_format", &data_format));

    // Grad for A
    grad_inputs[0] = upstream_grad;
    grad_inputs[0]->Ref();

    return nullptr;
  } else {
    return ToOperation(iter->second);
  }
}

TF_Operation* TF_GraphNextOperation(TF_Graph* graph, size_t* pos) {
  if (*pos == 0) {
    // Advance past the first sentinel nodes in every graph (the source & sink).
    *pos += 2;
  } else {
    // Advance to the next node.
    *pos += 1;
  }

  mutex_lock l(graph->mu);

    return tensorflow::errors::Internal("Graph pass runs for more than once!");
  }

 private:
  bool first_call_ = true;
};

// After the graph pass is registered, it takes effect globally and can affect
// other test cases. Define a static variable to switch it on and off.
bool GraphErrorInjectionPass::enabled_ = false;

// Test to ensure that a registered graph optimization pass is only executed

                           &incorrect_components, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_INVALID_ARGUMENT)
        << TF_Message(status.get());
  }

  {
    // Try to pass a ParallelTensor to TPUReplicatedInput
    std::array<TFE_TensorHandle*, 1> correct_components{value_one.get()};
    TensorHandlePtr combined_value = CreatePerDeviceValues(

      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }

  AbstractContextPtr immediate_execution_ctx_;
  GradientRegistry registry_;
  Status status_;

// This builder is used in the eager API to build a NodeDef.
struct TF_AttrBuilder : public tensorflow::AttrBuilder {
  using tensorflow::AttrBuilder::AttrBuilder;
  // The string buffers to make sure that any `attr_name` we pass into
  // `builder->Set()` will outlive the subsequent
  // `TF_AttrBuilderCheckCanRunOnDevice()` call(s) on the same `builder`.
  std::set<std::string> attr_names;
};

          std::unique_ptr<TensorSliceReader::VarToDataTypeMap>>
CheckpointReader::BuildV2VarMaps() {
  CHECK(v2_reader_ != nullptr);
  CHECK(v2_reader_->status().ok());

  // First pass: filters out the entries of the slices.
  std::unordered_set<string> filtered_keys;
  BundleEntryProto entry;
  v2_reader_->Seek(kHeaderEntryKey);
  for (v2_reader_->Next(); v2_reader_->Valid(); v2_reader_->Next()) {

      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }

  AbstractContextPtr immediate_execution_ctx_;
  GradientRegistry registry_;
  Status status_;

  if (n == 0) {
    TF_SetStatus(status, TF_OK, "");
    return 0;
  }
  if (!IsCacheEnabled() || (n > max_bytes_)) {
    // The cache is effectively disabled, so we pass the read through to the
    // fetcher without breaking it up into blocks.
    return block_fetcher_(filename, offset, n, buffer, status);
  }
  // Calculate the block-aligned start and end of the read.

      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }

  AbstractContextPtr immediate_execution_ctx_;
  GradientRegistry registry_;
  Status status_;