auto num_dims_analytical = TF_NumDims(analytical_tensor);
  ASSERT_EQ(dims.size(), num_dims_analytical);
  for (int j = 0; j < num_dims_analytical; j++) {
    auto dim_analytical = TF_Dim(analytical_tensor, j);
    ASSERT_EQ(dims[j], dim_analytical);
    num_elem_analytical *= dim_analytical;
  }

  float* danalytical = new float[num_elem_analytical]{0};
  memcpy(&danalytical[0], TF_TensorData(analytical_tensor),

        "Tensor expected to have rank 2 found rank: ", shape.dims());
  }
  int64_t dim_sizes[] = {2, 4};
  for (int i = 0; i < shape.dims(); i++) {
    if (shape.dim_size(i) != dim_sizes[i]) {
      return errors::InvalidArgument("Dim ", i, " expected to be of size ",
                                     dim_sizes[i],
                                     " found: ", shape.dim_size(i));
    }
  }
  return absl::OkStatus();
}

  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.
  virtual Status Dim(int dim_index, int64_t* dim) const = 0;

  // Returns the device which created the handle.
  virtual const char* DeviceName(Status* status) const = 0;
  // Returns the device where the tensor was placed.

  }
  auto imm_ctx = dyn_cast<ImmediateExecutionContext>(ctx);
  AbstractTensorPtr minus_1(imm_ctx->CreateInt32Scalar(-1));
  ImmediateTensorHandlePtr dim(imm_ctx->CreateLocalHandle(minus_1.get()));
  AbstractTensorHandle* expand_dims_outputs;
  TF_RETURN_IF_ERROR(
      ops::ExpandDims(ctx, vec, dim.get(), &expand_dims_outputs, "ExpandDims"));
  TF_RETURN_IF_ERROR(
      ops::Mul(ctx, expand_dims_outputs, mat, &outputs[0], "Mul"));

}

int64_t TFE_TensorHandleDim(TFE_TensorHandle* h, int dim_index,
                            TF_Status* status) {
  if (h == nullptr) {
    status->status = tensorflow::errors::InvalidArgument("Invalid handle");
    return -1;
  }

  int64_t dim = -1;
  status->status = tensorflow::unwrap(h)->Dim(dim_index, &dim);
  return dim;
}

          int64_t axis_length = combined_shape.dim_size(axis_index);
          if (axis_length != component_shape.dim_size(axis_index)) {
            axis_length = -1;
          }
          TF_RETURN_IF_ERROR(
              combined_shape.SetDimWithStatus(axis_index, axis_length));
        }
      }
    }
    auto dim_sizes = combined_shape.dim_sizes();
    shape_ = std::vector<int64_t>(dim_sizes.begin(), dim_sizes.end());
  }

    return nullptr;
  }
  tensorflow::PartialTensorShape partial_shape;
  if (shape.num_dims != -1) {
    DCHECK(shape.dim_sizes != nullptr);
    Status status = tensorflow::PartialTensorShape::MakePartialShape(
        reinterpret_cast<int64_t*>(shape.dim_sizes), shape.num_dims,
        &partial_shape);
    if (!status.ok()) {
      tsl::Set_TF_Status_from_Status(s, status);
      return nullptr;
    }

TF_CAPI_EXPORT extern int TFE_TensorDebugInfoOnDeviceNumDims(
    TFE_TensorDebugInfo* debug_info);

// Returns the number of elements in dimension `dim_index`.
// Tensor representation on device can be transposed from its representation
// on host. The data contained in dimension `dim_index` on device
// can correspond to the data contained in another dimension in on-host
// representation. The dimensions are indexed using the standard TensorFlow

    const int64_t* dims) {
  if (num_dims != -1) {
    std::vector<tensorflow::shape_inference::DimensionHandle> dim_vec;
    dim_vec.reserve(num_dims);
    for (int i = 0; i < num_dims; ++i) {
      dim_vec.push_back(ic->MakeDim(dims[i]));
    }
    return ic->MakeShape(dim_vec);
  } else {
    return ic->UnknownShape();
  }
}

}  // namespace

// dimension of a parallel tensor.
int64_t ParallelTensorDim(void* data, int dim_index, TF_Status* status) {
  const std::vector<int64_t>* shape;
  Status s = reinterpret_cast<ParallelTensor*>(data)->Shape(&shape);
  if (!s.ok()) {
    tsl::Set_TF_Status_from_Status(status, s);
    return -1;
  }
  return (*shape)[dim_index];
}

TF_Buffer* ParallelTensorSummarize(void* data, TF_Status* status) {