# tensorplay.autograd.graph API

Source: https://www.tensorplay.cn/docs/api/tensorplay.autograd.graph.html

## Functions 2

[#](#api-tensorplay.autograd.graph.save_on_cpu)

### save_on_cpu

function[Full reference ↗](/docs/generated/tensorplay.autograd.graph.save_on_cpu.html)

```python
tensorplay.autograd.graph.save_on_cpu(pin_memory: bool = False, device_type: str = 'cuda')
```

Store saved values in host memory and restore their original device.

[#](#api-tensorplay.autograd.graph.saved_tensors_hooks)

### saved_tensors_hooks

function[Full reference ↗](/docs/generated/tensorplay.autograd.graph.saved_tensors_hooks.html)

```python
tensorplay.autograd.graph.saved_tensors_hooks(pack_hook, unpack_hook)
```

Context manager that installs a (pack, unpack) pair for every tensor
saved via Function()’s save_for_backward
while active.

- pack_hook(tensor) runs once per saved tensor at save time and may return any object (e.g. a CPU copy, a key into external storage).

- unpack_hook(packed) runs when ctx.saved_tensors reads it back and must return an equivalent [tensorplay.Tensor](/docs/generated/tensorplay.Tensor.html#tensorplay.Tensor).

If either hook raises, the exception is surfaced from the corresponding
operation.