* Fixed issue when using OpenStack config drive for node metadata. Since we need to run commands such as blkid, we need to ensure that api server and kube controller are running in the privileged mode. ([#57561](https://github.com/kubernetes/kubernetes/pull/57561), [@dims](https://github.com/dims))

  TestHandleFromDLPack(status, ctx, {4}, {});
  TestHandleFromDLPack(status, ctx, {4}, {1});
  TestHandleFromDLPack(status, ctx, {4, 3, 2}, {});
  TestHandleFromDLPack(status, ctx, {4, 3, 2}, {6, 2, 1});
  // Test that dims with size=1 can have any stride.
  TestHandleFromDLPack(status, ctx, {1}, {1});
  TestHandleFromDLPack(status, ctx, {1}, {0});
  TestHandleFromDLPack(status, ctx, {4, 1, 2}, {2, 1, 1});

* Fix issue where on re-registration of device plugin, `allocatable` was not getting updated. This issue makes devices invisible to the Kubelet if device plugin restarts. Only work-around, if this fix is not there, is to restart...

- Update to latest cAdvisor 0.55.0 in our vendor dependencies ([#135829](https://github.com/kubernetes/kubernetes/pull/135829), [@dims](https://github.com/dims)) [SIG Node]
- Using pytorch based e2e integration test instead of tensorflow in some node e2e CI tests. ([#136397](https://github.com/kubernetes/kubernetes/pull/136397), [@dims](https://github.com/dims)) [SIG Testing]

    size_t size = TFE_TensorHandleDeviceMemorySize(copy, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    int64_t dims[] = {2, 2};
    TFE_TensorHandle* copy_aliased = TFE_NewTensorHandleFromDeviceMemory(
        ctx, name, TF_FLOAT, dims, 2, data, size, &Deleter, copy, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TFE_TensorHandle* on_host =

  // `memory_releaser` will be called on destruction, and it's responsible for
  // cleaning up the underlying buffer.
  virtual AbstractTensorInterface* CreateTensor(
      DataType dtype, const int64_t* dims, int num_dims, void* data, size_t len,
      MemoryReleaser memory_releaser, void* memory_releaser_arg) = 0;

  // Create a handle to wrap and manage a Tensor
  virtual ImmediateExecutionTensorHandle* CreateLocalHandle(

}

TF_Tensor* TFE_AllocateHostTensor(TFE_Context* ctx, TF_DataType dtype,
                                  const int64_t* dims, int num_dims,
                                  TF_Status* status) {
  std::vector<int64_t> dimvec(num_dims);
  for (int i = 0; i < num_dims; ++i) {
    dimvec[i] = static_cast<int64_t>(dims[i]);
  }

  if (ctx == nullptr) {
    status->status = tensorflow::errors::InvalidArgument("Invalid Context");

* kubeadm: Fix pause image to not use architecture, as it is a manifest list ([#65920](https://github.com/kubernetes/kubernetes/pull/65920), [@dims](https://github.com/dims))
* bazel deb package bugfix: The kubeadm deb package now reloads the kubelet after installation ([#65554](https://github.com/kubernetes/kubernetes/pull/65554), [@rdodev](https://github.com/rdodev))

  TFE_DeleteContextOptions(opts);

  std::vector<int64_t> dims(4, 1);
  TFE_Op* op = TFE_NewOp(ctx, "AvgPool", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TF_Tensor* tensor =
      TF_AllocateTensor(TF_FLOAT, dims.data(), dims.size(), sizeof(float));
  float tensor_data[] = {1};

//
// Custom devices must use TFE_NewCustomDeviceTensorHandle instead.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandleFromDeviceMemory(
    TFE_Context* ctx, const char* device_name, TF_DataType, const int64_t* dims,
    int num_dims, void* data, size_t len,
    void (*deallocator)(void* data, size_t len, void* arg),
    void* deallocator_arg, TF_Status* status);