To install the current release, which includes support for
[CUDA-enabled GPU cards](https://www.tensorflow.org/install/gpu) *(Ubuntu and
Windows)*:

```
$ pip install tensorflow
```

Other devices (DirectX and MacOS-metal) are supported using
[Device plugins](https://www.tensorflow.org/install/gpu_plugins#available_devices).

A smaller CPU-only package is also available:

```
$ pip install tensorflow-cpu
```

  let summary = "Lifting resource operations out of device computation";
  let description = [{
    This pass lifts resource variable operations outside of device computation.
    This is useful because a lot of accelerator devices can not interact with
    resource variables directly..

    Here is a simple example in TensorFlow where a device doubles the value of a
    TensorFlow resource variable and returns new value:

    ```mlir

The final graph looks like below:

![Final graph](g3doc/images/final_graph.png)

## TAC components

### Hardwares

Hardwares are used to modeling target device capabilities & also ops cost for
the target devices.

We have already modeled `cpu_hardware` & `gpu_hardware` as well as the
`nnapi_hardware`.

### Passes

#### Target Annotation Pass

  /// co-located on the device where op is placed.
  /// NOTE: This function is intended to be use internal libraries only for
  /// controlling placement of ops on to devices. Public use is not encouraged
  /// because the implementation of device placement is subject to change.
  Scope ColocateWith(const Operation& op) const;
  /// Convenience function for above.

			rcInstance.populatorHasAddedPods = func() bool {
				// Mark DSW populated to allow unmounting of volumes.
				return true
			}
			// Mark devices paths as reconciled to allow unmounting of volumes.
			rcInstance.volumesNeedUpdateFromNodeStatus = nil

			// Act 2 - reconcile once
			rcInstance.reconcile()

			// Assert 2
			// MountDevice was attempted

						},
					},
					CDIDevices: map[string][]string{
						"test-driver.cdi.k8s.io": {"example.com/example=cdi-example"},
					},
				},
			},
		},
		{
			"Restore checkpoint - single claim - multiple devices",

  }

  void getDependentDialects(DialectRegistry &registry) const override {
    registry.insert<TFL::TensorFlowLiteDialect>();
  }

 private:
  // The mode of TF Quantization, might indicate different users/devices.
  Option<std::string> mode_{*this, "mode",
                            llvm::cl::desc("The mode of TF Quantization."),
                            llvm::cl::init("")};

type Metrics struct {
	// The time at which these stats were updated.
	Time metav1.Time

	// Used represents the total bytes used by the Volume.
	// Note: For block devices this maybe more than the total size of the files.
	Used *resource.Quantity

	// Capacity represents the total capacity (bytes) of the volume's
	// underlying storage. For Volumes that share a filesystem with the host

                                 PatternRewriter &rewriter) const {
    return true;
  }

  // Class users should override this method if there are any op-specific
  // compatibility requirements for devices.
  virtual bool IsDeviceCompatible(SrcOpT contraction_op, BiasAddOp bias_add,
                                  PatternRewriter &rewriter) const {
    return true;
  }

        if (n->IsConstant()) {
          // We must decluster Const nodes that have an input control edge from
          // a different device, because this node may be part of the
          // co-ordination of while loops between devices.
          for (auto it : n->in_edges()) {
            if (!it->src()->assigned_device_name().empty() &&
                it->src()->assigned_device_name() !=
                    n->assigned_device_name()) {