PyTorch compiler and checkpoint proposals

PyTorch developers filed a cluster of compiler and checkpointing proposals, led by an Inductor change for non-Triton fused kernels. The rest are early issues on higher-order ops, debug logging, and custom operators, plus one DCP load optimization already under automated review.

Inductor fx wrapper for non-Triton fused kernels

A proposal would extend Inductor's fx_wrapper beyond Triton kernels so it can also host non-Triton fused kernels, including CPU fused ops. The change is an opening proposal, not a landed patch, and would widen which fused kernels the FX wrapper can represent.

Batched async prefetch for DCP FsspecReader

A pull request adds batched cat_ranges and async prefetching to FsspecReader, aimed at faster DCP loads. Automated review has flagged blocking issues. The thread has seven messages from five participants.

Higher-order ops, torch.compile, and dispatch subclasses

A placeholder issue was opened to explore making higher-order ops work with torch_dispatch subclasses and torch.compile. It is a starting point for that combination, not a design or implementation.

Side effects in higher-order ops

A single issue opens discussion of supporting side effects in PyTorch higher-order ops, with checkpointing as the stated use. No implementation is attached.

Debug mode support for pipeline parallelism

A single issue asks for debug logging order support under pipeline parallelism, citing AutoParallel barriers. It is a request to extend debug mode, not a completed change.

Associating functional custom ops with in-place variants

An opening issue proposes a way to associate functional custom operators with their in-place out variants. The stated reason is so Inductor memory planning can use that link.