PrivateUse1
A reserved device-type backend identifier in PyTorch that allows vendors to register custom accelerators as this type. Vendors need only implement aten operator dispatch and device guards to run most models.
Timeline (last 90 days)
PyTorch reserves a private namespace called PrivateUse1 in its device type enumeration, allowing hardware vendors to bind their own devices to that namespace through a registration mechanism without forking or modifying the PyTorch main codebase, thereby obtaining a real DeviceType.
When a vendor registers a device through PrivateUse1, it needs to register a device guard (DeviceGuard), a backend registry (dispatch key), a memory allocator, and several Python-side auxiliary interfaces.
After registration is complete, users can directly specify the hardware with a string such as torch.device("spyre"), and PyTorch maps it internally to PrivateUse1.
PyTorch reserves the PrivateUse1 device type, which vendors can register as their own accelerator; they only need to implement aten operator dispatch and device guards to run most models.
PyTorch extends hardware backends through the torch.device abstraction, the aten operator library, and the PyTorch Extensible Backend (torch_xpu / PrivateUse1) mechanism, allowing third-party chip vendors to integrate as plugins without modifying the main codebase.
All Facts (5)
PyTorch reserves a private namespace called PrivateUse1 in its device type enumeration, allowing hardware vendors to bind their own devices to that namespace through a registration mechanism without forking or modifying the PyTorch main codebase, thereby obtaining a real DeviceType.
50%UnverifiedWhen a vendor registers a device through PrivateUse1, it needs to register a device guard (DeviceGuard), a backend registry (dispatch key), a memory allocator, and several Python-side auxiliary interfaces.
50%UnverifiedAfter registration is complete, users can directly specify the hardware with a string such as torch.device("spyre"), and PyTorch maps it internally to PrivateUse1.
50%UnverifiedPyTorch reserves the PrivateUse1 device type, which vendors can register as their own accelerator; they only need to implement aten operator dispatch and device guards to run most models.
50%UnverifiedPyTorch extends hardware backends through the torch.device abstraction, the aten operator library, and the PyTorch Extensible Backend (torch_xpu / PrivateUse1) mechanism, allowing third-party chip vendors to integrate as plugins without modifying the main codebase.
50%