// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output
// Tensors need gradients to be computed (Tensors which are not used do not need
// any gradients to be computed).
//
// Finally, we start a backprop stack with a set of tape entries for which we
// have all gradients available. This set usually is a subset of the set of

  TF_DeleteBuffer(op_list_buf);
}

class DummyKernel : public tensorflow::OpKernel {
 public:
  explicit DummyKernel(tensorflow::OpKernelConstruction* context)
      : OpKernel(context) {}
  void Compute(tensorflow::OpKernelContext* context) override {}
};

// Test we can query kernels
REGISTER_OP("TestOpWithSingleKernel")
    .Input("a: float")
    .Input("b: float")
    .Output("o: float");

  TF_Status* s = TF_NewStatus();
  TF_OpKernelConstruction_GetAttrType(ctx, "foobar", &type, s);
  TF_DeleteStatus(s);
  return NULL;
}

// A compute function. This will never actually get called in this test, it's
// just nice to know that it compiles.
void compute(void* kernel, TF_OpKernelContext* ctx) {
  TF_Tensor* input;
  TF_Status* s = TF_NewStatus();
  TF_GetInput(ctx, 0, &input, s);
  TF_DeleteTensor(input);

*   TF Core:

    *   Corrected higher-order gradients of control flow constructs (`tf.cond`,
        `tf.while_loop`, and compositions like `tf.foldl`) computed with
        `tf.GradientTape` inside a `tf.function`.
    *   Changed the default step size in `gradient_checker_v2.compute_gradients`
        to be exactly representable as a binary floating point numbers. This

*   When you contribute a new feature to TensorFlow, the maintenance burden is
    (by default) transferred to the TensorFlow team. This means that the benefit
    of the contribution must be compared against the cost of maintaining the
    feature.
*   Full new features (e.g., a new op implementing a cutting-edge algorithm)
    typically will live in

# with clang. It does not imply that CUDA support has been enabled.
alias(
    name = "is_cuda_compiler_clang",
    actual = if_oss(
        "@local_config_cuda//:is_cuda_compiler_clang",
        "@local_config_cuda//cuda:TRUE",
    ),
)

# Config setting that is satisfied when CUDA device code should be compiled

  buf->Unref();
  return tensorflow::wrap(tensorflow::TensorHandle::CreateLocalHandle(
      std::move(t), device, device, context));
}

// This function will block till the operation that produces `h` has
// completed. This is only valid on local TFE_TensorHandles. Returns the size in
// bytes of the memory pointed to by the device pointer returned above.
size_t TFE_TensorHandleDeviceMemorySize(TFE_TensorHandle* h,

build:cuda_clang --@local_config_cuda//:cuda_compiler=clang
# Select supported compute capabilities (supported graphics cards).
# This is the same as the official TensorFlow builds.
# See https://developer.nvidia.com/cuda-gpus#compute
# `compute_XY` enables PTX embedding in addition to SASS. PTX
# is forward compatible beyond the current compute capability major
# release while SASS is only forward compatible inside the current

  return output


def set_tf_cuda_compute_capabilities(environ_cp):
  """Set TF_CUDA_COMPUTE_CAPABILITIES."""
  while True:
    native_cuda_compute_capabilities = get_native_cuda_compute_capabilities(
        environ_cp)
    if not native_cuda_compute_capabilities:
      default_cuda_compute_capabilities = _DEFAULT_CUDA_COMPUTE_CAPABILITIES
    else:
      default_cuda_compute_capabilities = native_cuda_compute_capabilities

      control_output_names_vec.push_back(string(control_output_names[i]));
    }
  }

  // Compute body nodes.
  std::vector<const Node*> body_nodes;
  status->status = tensorflow::ComputeBodyNodes(
      fn_body, fn_name, num_opers, opers, input_nodes, &body_nodes);
  if (TF_GetCode(status) != TF_OK) return nullptr;

  // Compute body nodes.
  std::vector<const Node*> control_output_nodes;