Track the torch-infini integration roadmap

[voltjia]

Goal

Advance torch-infini from the initial PrivateUse1 bridge to a reproducible, stream-correct InfiniOps-backed PyTorch plugin while keeping each pull request focused on one responsibility.

Ordered work

Runtime correctness

• Prefer an available InfiniRT accelerator backend #2 Prefer an available InfiniRT accelerator backend.
• Preserve current device across runtime operations #3 Preserve current device across explicit-device operations.
• Implement the InfiniRT stream bridge #6 Implement native stream lifecycle and current-stream semantics.
• Add the public InfiniRT event bridge #7 Add the public InfiniRT event bridge.
• Make allocation and copy stream-aware #10 Make allocation and copy stream-aware.

Build and validation

• Add GitHub-hosted CPU build and test CI #4 Add GitHub-hosted CPU build/test CI.
• Make wheels independent of the InfiniRT build prefix #5 Remove build-prefix RUNPATH from wheels.
• Add trusted NVIDIA build and test CI #11 Add trusted self-hosted NVIDIA build/test CI.
• Define the PyTorch and InfiniRT compatibility matrix #12 Define and test the PyTorch/InfiniRT compatibility matrix.

Operator integration

• Add InfiniOps build integration and tensor adapters #8 Add InfiniOps discovery, linking, and tensor adapters.
• Implement aten::add.Tensor through InfiniOps #9 Implement aten::add.Tensor through InfiniOps.

Parallelism

After #2, #3, #4, #5, #8, and the NVIDIA CI infrastructure can progress independently. Stream work (#6) and the InfiniOps adapter (#8) can also proceed in parallel. The first computational operator (#9) waits for both. Public events and async copy remain separate follow-ups so stream support does not silently expand into allocator policy.

Later operator generation, autograd, AMP, RNG, distributed support, and release wheel policy should be opened only after the first native operator path and CI matrix provide a tested baseline.