<< type_string() << " on an XLA device. This should never happen.";
}

XlaAssignVariableOp::XlaAssignVariableOp(OpKernelConstruction* c)
    : OpKernel(c) {
  OP_REQUIRES_OK(c, c->GetAttr("dtype", &dtype_));
}

void XlaAssignVariableOp::Compute(OpKernelContext* context) {
  OP_REQUIRES(context, dtype_ == context->input(1).dtype(),
              errors::InvalidArgument(

    }
  }

  if (!in_order) {
    // Functions do not get reused in practice, so skip the check for if the
    // callee has been updated.
    StringAttr callee_sym = cluster_func_op.getFuncAttr().getAttr();
    MoveResourceArgsToEnd(symtab.lookup<func::FuncOp>(callee_sym));
  }
  builder.setInsertionPoint(cluster_func_op);
  auto xla_launch_op = builder.create<TF::XlaLaunchOp>(

  StringAttr device = mlir::cast<StringAttr>(
      added_func_op->getRegion(0).getBlocks().front().front().getAttr(kDevice));
  StringAttr inference_type = mlir::cast<StringAttr>(
      added_func_op->getRegion(0).getBlocks().front().front().getAttr(
          kInferenceType));
  added_call_op->setAttr(kDevice, device);
  added_call_op->setAttr(kInferenceType, inference_type);

	a.Atr = attr
	a.Cls = uint8(cls)
	a.Value = value
	a.Data = data
}

// Each DIE (except the root ones) has at least 1 attribute: its
// name. getattr moves the desired one to the front so
// frequently searched ones are found faster.
func getattr(die *dwarf.DWDie, attr uint16) *dwarf.DWAttr {
	if die.Attr.Atr == attr {
		return die.Attr
	}

	a := die.Attr
	b := a.Link
	for b != nil {

  LogicalResult matchAndRewrite(mlir::stablehlo::CustomCallOp op,
                                PatternRewriter &rewriter) const override {
    auto backendConfig =
        mlir::dyn_cast<DictionaryAttr>(op->getAttr("composite.backend_config"));
    if (!backendConfig)
      return op->emitError(
          "custom_call has no 'composite.backend_config' attribute or the "
          "attribute is not a dictionary");

                                PatternRewriter& rewriter) const override {
    auto input_op = op.getArg().getDefiningOp();
    if (auto q = llvm::dyn_cast_or_null<quantfork::QuantizeCastOp>(input_op)) {
      if (!q->getAttr(kVolatileOpAttrName)) return failure();

      if (remove_volatile_ops_type == kPreserveInputsAndOutputs) {
        // Don't remove leading and trailing QDQ for PTQ workflow, so the io
        // modifying lib can work correctly.

    if (op->getDialect() != tf_dialect) return WalkResult::advance();

    if (parallel_group_attr) {
      op->setAttr(TF::kParallelExecAnnotation, parallel_group_attr);
    }
    auto device_attr = op->getAttr(kDeviceAttr);
    if (!device_attr) {
      op->setAttr(kDeviceAttr, launch.getDeviceAttr());
      return WalkResult::advance();
    }

    if (auto device_str_attr = mlir::dyn_cast<StringAttr>(device_attr)) {

    if operation_id in operation_ids:
        message = (
            f"Duplicate Operation ID {operation_id} for function "
            + f"{route.endpoint.__name__}"
        )
        file_name = getattr(route.endpoint, "__globals__", {}).get("__file__")
        if file_name:
            message += f" at {file_name}"
        warnings.warn(message, stacklevel=1)
    operation_ids.add(operation_id)

    auto cond = symbol_table.lookup<mlir::func::FuncOp>(
        while_op->getAttr("cond").cast<mlir::FlatSymbolRefAttr>().getValue());
    AddCallOpInWhileOpRegion(while_op.getCond(), cond);
    while_op->removeAttr("cond");
    auto body = symbol_table.lookup<mlir::func::FuncOp>(
        while_op->getAttr("body").cast<mlir::FlatSymbolRefAttr>().getValue());
    AddCallOpInWhileOpRegion(while_op.getBody(), body);

                                PatternRewriter& rewriter) const override {
    auto input_op = op.getArg().getDefiningOp();
    if (auto q = llvm::dyn_cast_or_null<quantfork::QuantizeCastOp>(input_op)) {
      if (!q->getAttr(kVolatileOpAttrName)) return failure();

      // If the quantize op is a requantize op, it is being used in other scale
      // adjustments and should be kept. Instead, move dequantize op before the