// Enables running eager ops as functions.
  virtual void SetRunEagerOpAsFunction(bool enable) = 0;

  // Enables rewriting jit_compile functions.
  virtual void SetJitCompileRewrite(bool enable) = 0;

  // Sets the device placement policy for the current thread.
  virtual void SetThreadLocalDevicePlacementPolicy(
      ContextDevicePlacementPolicy policy) = 0;
  // Returns the device placement policy for the current thread.

// enabled through MarkForCompilationPassFlags.
TF_CAPI_EXPORT unsigned char TF_GetXlaAutoJitEnabled();

// Sets XLA's minimum cluster size. This has global effect.
TF_CAPI_EXPORT void TF_SetXlaMinClusterSize(int size);

// Gets/Sets TF/XLA flag for whether(true) or not(false) to disable constant
// folding. This is for testing to ensure that XLA is being tested rather than

    constraint_values = if_google(
        ["//third_party/bazel_platforms/os:fuchsia"],
        [],
    ),
    values = if_oss(
        # TODO(b/149248802) When we have a Fuchsia Bazel SDK update to use the values it sets.
        {"cpu": "fuchsia"},
        {},
    ),
    visibility = ["//visibility:public"],
)

config_setting(
    name = "fuchsia_x86_64",
    constraint_values = if_google(

#include "tensorflow/c/eager/abstract_context.h"
#include "tensorflow/c/eager/gradients.h"

namespace tensorflow {
namespace gradients {
// Ignores `grad_outputs` and sets all entries in grad_inputs to nullptr.
class NotDifferentiableGradientFunction : public GradientFunction {
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,

  /// Releases all resources used by the filesystem
  ///
  /// NOTE: TensorFlow does not unload DSOs. Thus, the only way a filesystem
  /// won't be registered anymore is if this function gets called by core
  /// TensorFlow and the `TF_Filesystem*` object is destroyed. However, due to
  /// registration being done in a static instance of `Env`, the destructor of
  /// `FileSystem` is never called (see

  TensorReference reference;
  std::vector<int64_t> shape;
  std::vector<int64_t> strides;
  DLManagedTensor tensor;

  explicit TfDlManagedTensorCtx(const TensorReference& ref) : reference(ref) {}
};

// Gets tensor from eager tensor handle.
const Tensor* GetTensorFromHandle(TFE_TensorHandle* h, TF_Status* status) {
  if (h == nullptr) {
    status->status = tensorflow::errors::InvalidArgument("Invalid handle");
    return nullptr;

    change.
-   You make the change and submit it for the review again.
-   This cycle repeats itself until the PR gets approved.
-   Note: As a friendly reminder, we may reach out to you if the PR is awaiting
    your response for more than 2 weeks.

**4. Approved**

-   Once the PR is approved, it gets `kokoro:force-run` label applied and it
    initiates CI/CD tests.
-   We can't move forward if these tests fail.

// `g` will be returned. Else, an unspecified subset.
//
// If successful, returns the number of TF_Function* successfully set in
// `funcs` and sets status to OK. The caller takes ownership of
// all the returned TF_Functions. They must be deleted with TF_DeleteFunction.
// On error, returns 0, sets status to the encountered error, and the contents
// of funcs will be undefined.

  ~ParallelDevice();

  // Helper to copy a tensor handle from another device once for each component
  // of the ParallelDevice.
  //
  // Sets a bad status and returns a nullptr if `tensor` is already on the
  // ParallelDevice, or if the individual copies fail.
  std::unique_ptr<ParallelTensor> CopyToParallelDevice(TFE_Context* context,

  //
  // Executes the backward function in order to trace its gradient, which will
  // waste computation if executing eagerly (when graph building the unneeded
  // computation is pruned). Temporarily sets `backward_tape` so that
  // Accumulate will forward op executions to the tape while the backward
  // function is running; this effectively adds the backward tape to the active