void MarkAsResult(AbstractTensorHandle* gradient) const override;

  void DeleteGradient(AbstractTensorHandle* gradient) const override;

 private:
  // The context where the aggregation op `Add` is to be created.
  AbstractContext* ctx_;
};

// Returns the number of elements in the gradient tensor.
int64_t TapeVSpace::NumElements(AbstractTensorHandle* tensor) const {

  }
  void runOnOperation() override;
};

// Creates a call to function `func` in region `caller_region`. Use `args` as
// the call arguments, and terminate the region with a yield. The arguments are
// cast to the required type before the call. `use_region_args` control whether
// the input arguments are used as is (for IfOp) or block arguments of the same
// type as the input arguments are created and then used as call arguments (for
// While).

      builder.clone(*op, mapping);
    }
    // Creates the ReturnOp so that the per-host function returns the
    // correct values of the cloned operations.
    llvm::SmallVector<Value, 4> results_after_mapping;
    for (Value result : metadata.results) {
      results_after_mapping.push_back(mapping.lookupOrDefault(result));
    }
    builder.create<func::ReturnOp>(loc, results_after_mapping);

// Caches a XlaDeviceAllocator per <backend, device ordinal> pair. A
// XlaDeviceAllocator is created on demand and is associated with a
// XlaDevice. It outlives the device itself (for instance, the buffer
// backing a tensor holds a pointer to the allocator for book-keeping,
// and this buffer can outlast the device).
class XlaDeviceAllocatorState {
 public:
  // Creates or returns a cached XlaDeviceAllocator for a given
  // backend and device_ordinal.

          !IsQintValueDefinedByIntToQintCast(op->getOperand(i))) {
        new_operands.push_back(rewriter.create<TF::CastOp>(
            op->getLoc(), orig_op_type, operands[i]));
      } else {
        new_operands.push_back(operands[i]);
      }
    }

    // Create a new UQ op.
    OperationState state(op->getLoc(), op->getName().getStringRef(),

  llvm::SmallVector<mlir::TF::SplitOp, 4> split_ops_for_tiled_input;
  split_ops_for_tiled_input.reserve(
      input_sharding.tile_assignment_devices_size());

  // Creates a tree of split nodes for sharding tiled inputs. Splits nodes
  // are created such that input data is sharded in row major order.
  // Split nodes at ith depth from the original input node represent nodes
  // that split the input data at i-th dimension.

//     probe for gadgets inside core TensorFlow. We can even protect the area
//     of memory where the copies reside to not allow any more writes to it
//     after all copies are created.
template <typename T>
static std::unique_ptr<const T> CopyToCore(const T* plugin_ops,
                                           size_t plugin_size) {
  if (plugin_ops == nullptr) return nullptr;

    };
    auto del_dealloc = [&](int pos) {
      deltas.emplace_back(pos + 1, tensor.size);
    };

    if (global.first_use < remat.begin) {
      // The tensor is created before the source range.
      if (global.last_use < remat.insert) {
        // It currently gets deallocated before the newly inserted range, so we
        // need to extend its lifetime: It will now be deallocated at its last

  builder.setInsertionPoint(&terminator);
  builder.create<tf_executor::YieldOp>(terminator.getLoc(),
                                       terminator.getOperands());
  terminator.erase();

  builder.setInsertionPoint(island_op);
  IRMapping mapping;
  for (int i : llvm::seq<int>(0, num_replicas)) {
    // Create new island for replica.
    auto replica = builder.create<tf_executor::IslandOp>(

      moved_ops_list.push_back(new_if_op.getOperation());
    }
  }
}

// Create the Yield ops for both branches with merged results.
// `builder` is the OpBuilder.
// `if_op_segment` refers to the segment of IfRegions to be merged.
// `return_indices` refers to the return indices to be kept in merged IfRegion
// `new_if_op` refers to the created new IfRegion
void CreateYieldOps(