(*var_to_data_type_map)[key] = DataType(entry.dtype());
  }
  // The returned pointers are owned by the caller.
  return std::make_pair(std::move(var_to_shape_map),
                        std::move(var_to_data_type_map));
}

}  // namespace checkpoint

  if (num_dims > TensorShape::MaxDimensions()) {
    return errors::InvalidArgument("Value specified for `", attr_name, "` has ",
                                   num_dims,
                                   " dimensions which is over the limit of ",
                                   TensorShape::MaxDimensions(), ".");
  }
  TensorShapeProto proto;
  if (num_dims < 0) {
    proto.set_unknown_rank(true);
  } else {

                               file.get(), plugin_status.get());

  if (TF_GetCode(plugin_status.get()) == TF_OK)
    *result = std::make_unique<ModularRandomAccessFile>(
        translated_name, std::move(file), random_access_file_ops_.get());

  return StatusFromTF_Status(plugin_status.get());
}

Status ModularFileSystem::NewWritableFile(
    const std::string& fname, TransactionToken* token,

to the actual TensorFlow wheel.
3. Creates a `requirements_lock_<python_version>.txt` file.
4. Updates the `requirements_lock_<python_version>.txt` file using
a Bazel command.
5. Moves the updated `requirements_lock_<python_version>.txt` file
to the `../wheel_test/` directory.


### How it Works in the Presubmit Job

`_requirements_lock` files will be generated by the presubmit job. A detailed

  // case placement information is lost when resolving the handle.
  //
  // If false, a caller might implement pretty-printing by resolving and
  // iterating over the resulting tensor. This may still be viable if resolving
  // the handle loses information, but `SummarizeValue` would be more precise.
  virtual bool PreferCustomSummarizer() const { return false; }

void RamFileBlockCache::Trim() {
  while (!lru_list_.empty() && cache_size_ > max_bytes_) {
    RemoveBlock(block_map_.find(lru_list_.back()));
  }
}

/// Move the block to the front of the LRU list if it isn't already there.
void RamFileBlockCache::UpdateLRU(const Key& key,
                                  const std::shared_ptr<Block>& block,

  TF_Tensor* t = TF_NewTensor(TF_FLOAT, dims, 2, values, num_bytes,
                              &Deallocator, &deallocator_called);

  TF_Tensor* o = TF_TensorMaybeMove(t);
  ASSERT_TRUE(o == nullptr);  // It is unsafe to move memory TF might not own.
  TF_DeleteTensor(t);
  EXPECT_TRUE(deallocator_called);
}

TEST(CAPI, LibraryLoadFunctions) {
  // TODO(b/73318067): Fix linking for the GPU test generated by the

  // this path.
  //
  // Note that some filesystem might require a different approach here, for
  // example they might require the root directory path to be in a special
  // format, etc. When we get there, we might decide to move this class to
  // `modular_filesystem_test.h` and extend the instantiation to also take as
  // argument an implementation for this method/a subclass factory (see

              "localhost:", tensorflow::testing::PickUnusedPortOrDie())});

  tensorflow::JobDef* job_def2 = cluster_def->add_job();
  job_def2->set_name("worker");

  // Copy over `host:port` at `task_index`
  for (auto task : cluster_server_def.cluster().job(0).tasks()) {
    if (task.first == task_index) {
      job_def2->mutable_tasks()->insert({task.first, task.second});
    }
  }

  Cleanup(Cleanup&& src)  // NOLINT
      : released_(src.is_released()), f_(src.release()) {}

  // Implicitly move-constructible from any compatible Cleanup<G>.
  // The source will be released as if src.release() were called.
  // A moved-from Cleanup can be safely destroyed or reassigned.
  template <typename G>
  Cleanup(Cleanup<G>&& src)  // NOLINT