//
// Use in tests to allow XLA to fallback to TF classic. This has global effect.
TF_CAPI_EXPORT unsigned char TF_SetXlaEnableLazyCompilation(
    unsigned char enable);
TF_CAPI_EXPORT unsigned char TF_SetTfXlaCpuGlobalJit(unsigned char enable);

// Sets XLA's auto jit mode according to the specified string, which is parsed
// as if passed in XLA_FLAGS. This has global effect.
TF_CAPI_EXPORT void TF_SetXlaAutoJitMode(const char* mode);

// in_flight_nodes_limit: when is_async is true, this value controls the
// maximum number of in flight async nodes. Enqueuing of additional async ops
// after the limit is reached blocks until some inflight nodes finishes.
// The effect is bounding the memory held by inflight TensorHandles that are
// referenced by the inflight nodes.
// A recommended value has not been established.
// A value of 0 removes the limit, which is the behavior of TensorFlow 2.11.

// an operation using the function's name.
// Any changes to `func`/`grad` (including deleting it) done after this method
// returns, won't affect the copy of `func`/`grad` in `g`.
// If `func` or `grad` are already in `g`, TF_GraphCopyFunction has no
// effect on them, but can establish the function->gradient relationship
// between them if `func` does not already have a gradient. If `func` already

// capture some inputs and then add a node in the graph. The output tensors are
// returned through the provided TF_OutputList.
// Any active tape will observe the effects of this execution.
void TF_ExecuteOperation(TF_AbstractOp* op, int num_inputs,
                         TF_AbstractTensor* const* inputs, TF_OutputList* o,
                         TF_Status* s);

  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  ASSERT_EQ(TF_TensorType(actual_value.get()),
            static_cast<TF_DataType>(DataTypeToEnum<value_type>().value));
  EXPECT_EQ(expected_value,
            *static_cast<value_type*>(TF_TensorData(actual_value.get())));
}

template <std::size_t num_devices>
void RegisterParallelDevice(
    TFE_Context* context, const char* device_name,

// Sets a thread-local device placement policy. After this call, other calls to
// TFE_Execute in the same thread will use the device policy specified here
// instead of the device policy used to construct the context. This has no
// effect on the device policy used by other program threads.
TF_CAPI_EXPORT extern void TFE_ContextSetThreadLocalDevicePlacementPolicy(
    TFE_Context* ctx, TFE_ContextDevicePlacementPolicy policy);

/// argument list with the out arguments. We only use the return type for an out
/// parameter if that is a plain C type, as this ensures ABI compatibility
/// (returning structures has issues in case compiler options affect
/// optimizations such as RVO). If a status needs to be returned from these
/// methods, the last argument is always a `TF_Status *` (or an array of such
/// pointers) owned by core TensorFlow and guaranteed to not be `nullptr`.