// used for compilation is output in `client`. The PJRT executable is output in
// `executable`.
Status CompileToPjRtLoadedExecutable(
    const OpKernelContext& ctx, const XlaPlatformInfo& platform_info,
    const NameAttrList& function,
    const std::vector<XlaCompiler::Argument>& args,
    DeviceCompileMode compile_mode, bool has_ref_vars,
    bool may_alias_resource_update,

      GetAllocator(ctx->device(), stream, platform_info_);
  se::DeviceMemoryAllocator* allocator = allocator_ptr.get();
  XlaComputationLaunchContext launch_context(
      client, allocator, client->default_device_ordinal(),
      /*allocate_xla_tensors=*/platform_info_.xla_device_metadata() != nullptr,
      platform_info_.xla_device_metadata()
          ? platform_info_.xla_device_metadata()->UseMultipleStreams()

  // Indexes of compile-time constant inputs
  const std::vector<int> constants_;
  // Indexes of resource inputs
  const std::vector<int> resources_;

  const NameAttrList function_;
  const XlaPlatformInfo platform_info_;

  bool has_ref_vars_;
};

// XlaLocalLaunchOp is used to replace a region of the TensorFlow graph
// which will be compiled and executed using XLA.  The XlaLocalLaunchOp is

// vanilla TensorFlow op as long as the bridge supports it.
class XlaCompileOnDemandOp : public OpKernel {
 public:
  explicit XlaCompileOnDemandOp(OpKernelConstruction* ctx)
      : OpKernel(ctx),
        platform_info_(XlaPlatformInfoFromDevice(ctx->device())) {}
  void Compute(OpKernelContext* ctx) override;

 private:
  Status Compile(const std::vector<XlaCompiler::Argument>& args,
                 OpKernelContext* ctx,