# DLPack

Source: https://www.tensorplay.cn/docs/dlpack.html

[DLPack](https://dmlc.github.io/dlpack/) is a neutral, zero-copy exchange format for tensors.
It lets two different libraries hand a tensor back and forth without copying the data,
describing only the shape, strides, data type, and device. TensorPlay interoperates with it
through tensorplay.from_dlpack.

```
import tensorplay as tp

# a DLPack capsule is an opaque handle to a tensor
x = tp.ones(3, 4)
capsule = tp.to_dlpack(x)
y = tp.from_dlpack(capsule)   # wraps the same data, no copy
print(y.shape)                # (3, 4)
```

tensorplay.from_dlpack wraps data owned by another library, and tensorplay.to_dlpack
hands ours out in the same format. Because the data is shared rather than copied, both ends
must respect the tensor’s lifetime and the device contract — the capsule owns the underlying
memory and the two sides agree on when it is valid.

| tensorplay.from_dlpack | |
| --- | --- |
| tensorplay.to_dlpack | Exports obj as a DLPack capsule without copying its data. |

## Where to go next

- [Tensors](/docs/tensorplay.html) — the tensor APIs from_dlpack and to_dlpack interoperate with.

- [NumPy interop](/docs/notes/serialization.html) — the other zero-copy conversion path.