std::vector<AbstractTensorHandle*> temp_outputs(2);
  TF_RETURN_IF_ERROR(
      ops::IdentityN(ctx, inputs, absl::MakeSpan(temp_outputs), "IdentityN"));
  // Although, `ops::IdentityN` returns 2 tensors, the first tensor isn't needed
  // for computing gradient so we could safely drop it.
  outputs[0] = temp_outputs[1];
  temp_outputs[0]->Unref();
  return absl::OkStatus();
}

class CppGradients

    AbstractTensorHandle* temp_output;

    // Calculate conjugate of X
    std::string name = "Conj_Log1p_Grad_X";
    TF_RETURN_IF_ERROR(SafeConj(ctx, X, &temp_output, name.c_str()));

    AbstractTensorHandlePtr Conj_X(temp_output);

    // Creates Ones
    name = "OnesLike_Log1p_Grad_X";
    TF_RETURN_IF_ERROR(OnesLike(ctx, Conj_X.get(), &temp_output, name.c_str()));

    large_constant]`) into 32-bit arithmetic in SASS.

    As a result, these versions of `ptxas` miscompile most XLA programs which
    use more than 4GB of temp memory. This results in garbage results and/or
    `CUDA_ERROR_ILLEGAL_ADDRESS` failures.

    A fix in CUDA 9.1.121 is expected in late February 2018. We do not expect a

      tape.Watch(inputs[i]);
    }
    std::vector<AbstractTensorHandle*> temp_outputs(1);
    AbstractContextPtr tape_ctx(new TapeContext(ctx, &tape, grad_registry));
    TF_RETURN_IF_ERROR(
        forward_model(tape_ctx.get(), inputs, absl::MakeSpan(temp_outputs)));

    TF_RETURN_IF_ERROR(tape.ComputeGradient(ctx, /*targets=*/temp_outputs,
                                            /*sources=*/inputs,