*    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();

      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_;

      ASSERT_EQ(errors::OK, s.code()) << s.message();
      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 ctx_;

 public:

          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_;