PyTorch CUDA kernels, Inductor RNG, and llama.cpp Gemma work

PyTorch pull requests dominated the day with new CUDA histogram kernels, Inductor RNG alignment, distributed checkpoint I/O work, and a small MPS refactor. llama.cpp also advanced graph optimizations aimed at Gemma models.

CUDA histogram and histogramdd kernels

A pytorch/pytorch pull request adds CUDA kernels for histogram and histogramdd, along with shared validation helpers and AOTI gating. Automated review judged the change ready. GPU users who need histogram operations obtain native CUDA paths that can reduce reliance on slower fallbacks for statistical workloads.

Gemma graph optimizations in llama.cpp

A ggml-org/llama.cpp pull request introduces graph optimizations for Gemma models, with the subject citing gains on prompt processing for E2B and E4B variants. A single comment flags the change. Operators running those Gemma configurations in llama.cpp stand to see improved prompt-processing throughput from the graph-level work.

Opt-in compiled CUDA RNG matching in Inductor

A pytorch/pytorch proposal makes opt-in compiled CUDA RNG match eager bitwise behavior and use all four Philox words. The change targets consistency between Inductor-compiled and eager execution. Developers who require bit-identical random streams across those paths gain a clearer opt-in route.

Batched async prefetch for DCP FsspecReader

A pytorch/pytorch pull request optimizes FsspecReader with batched cat_ranges and async prefetching to speed distributed checkpoint loads. Multiple participants discussed the change while automated review flagged blocking issues. Teams loading large checkpoints through fsspec backends have a concrete path to lower I/O latency once those blockers clear.

MPS check refactor for out-of-tree backends

A minor pytorch/pytorch pull request refactors hardcoded MPS checks in cpp_wrapper to better support out-of-tree backends. The edit reduces special-casing that complicated external device integration. Maintainers of alternative backends obtain cleaner entry points into the C++ wrapper path.