torch.utils.rename_privateuse1_backend — PyTorch 2.7 documentation (original) (raw)

torch.utils.rename_privateuse1_backend(backend_name)[source][source]

Rename the privateuse1 backend device to make it more convenient to use as a device name within PyTorch APIs.

The steps are:

  1. (In C++) implement kernels for various torch operations, and register them to the PrivateUse1 dispatch key.
  2. (In python) call torch.utils.rename_privateuse1_backend(“foo”)

You can now use “foo” as an ordinary device string in python.

Note: this API can only be called once per process. Attempting to change the external backend after it’s already been set will result in an error.

Note(AMP): If you want to support AMP on your device, you can register a custom backend module. The backend must register a custom backend module with torch._register_device_module("foo", BackendModule). BackendModule needs to have the following API’s:

  1. get_amp_supported_dtype() -> List[torch.dtype]get the supported dtypes on your “foo” device in AMP, maybe the “foo” device supports one more dtype.

Note(random): If you want to support to set seed for your device, BackendModule needs to have the following API’s:

  1. _is_in_bad_fork() -> boolReturn True if now it is in bad_fork, else return False.
  2. manual_seed_all(seed int) -> NoneSets the seed for generating random numbers for your devices.
  3. device_count() -> intReturns the number of “foo”s available.
  4. get_rng_state(device: Union[int, str, torch.device] = 'foo') -> TensorReturns a list of ByteTensor representing the random number states of all devices.
  5. set_rng_state(new_state: Tensor, device: Union[int, str, torch.device] = 'foo') -> NoneSets the random number generator state of the specified “foo” device.

And there are some common funcs:

  1. is_available() -> boolReturns a bool indicating if “foo” is currently available.
  2. current_device() -> intReturns the index of a currently selected device.

For more details, see https://pytorch.org/tutorials/advanced/extend_dispatcher.html#get-a-dispatch-key-for-your-backendFor an existing example, see https://github.com/bdhirsh/pytorch_open_registration_example

Example:

torch.utils.rename_privateuse1_backend("foo")

This will work, assuming that you've implemented the right C++ kernels

to implement torch.ones.

a = torch.ones(2, device="foo")