Motivation

OpenReg is PyTorch's in-tree reference implementation for PrivateUse1 backend integration. It currently covers operator registration, autoload, AMP, and device APIs but has no torch.compile integration. This means there is no reference how a PrivateUse1 device integrates with PyTorch's compiler stack.

Design

The concepts at the torch.compile interface are shared across all backends registration, FX graph capture, FakeTensor metadata, guards, graph breaks. But the actual compilation logic is completely different from one backend to the next (TPU lowers to StableHLO/XLA, CUDA targets Triton, etc.).

Following OpenReg's minimality principle, the proposed backend is a passthrough, it receives the FX graph from Dynamo, performs lightweight validation using FX APIs, and returns it for eager execution through the existing CPU fallback. No lowering, optimization, or device-specific code generation is performed.

The backend validates the PyTorch side integration contract, not backend-specific compilation logic. Each real vendor validates their own lowering in their own out-of-tree CI.

What the backend does

A function implementing the standard torch.compile backend contract:

• Validates graph well-formedness via Graph.lint()
• Checks FakeTensor device metadata on placeholder nodes to verify PrivateUse1 device propagation
• Demonstrates the transform-and-recompile pattern via eliminate_dead_code() and recompile()
• Logs generated source via python_code() for debug visibility
• Returns gm.forward (passthrough execution)
• Registered via in-code register_backend call, triggered on package import.

What the tests cover

Tests validating the PyTorch-side behavior for PrivateUse1:

• Backend registration and discoverability
• End-to-end compile and execute on PrivateUse1 tensors
• FakeTensor device propagation through Dynamo tracing
• Guard caching and shape-triggered recompilation
• Autograd through compilation
• nn.Module with PrivateUse1 parameters
• Dynamic shapes
• Graph break detection on device transfers
• Autocast with PrivateUse1 device type
• Default device interaction with compilation

Pull Request: 181092

cc @bdhirsh @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @amjames @Lucaskabela @jataylo @azahed98 @NmomoN @mengpenghui @fwenguang @cdzhan @1274085042 @PHLens @albanD