AI tooling: llama.cpp crash and PyTorch PRs

Security follow-ups and backend API work marked the day in AI and ML libraries. A known crash report resurfaced in llama.cpp while PyTorch saw two feature proposals.

Llama.cpp still crashes on nested JSON schemas

A report in the ggml-org/llama.cpp project states that CVE-2026-52130 continues to crash llama-server through unbounded nesting in the json-schema-to-grammar path. The observation was recorded at commit 19e28a277 under issue 29690. Operators who rely on schema-constrained generation should treat the defect as still open.

Dynamo registration APIs for out-of-tree backends

A pull request in pytorch/pytorch adds three registration APIs meant to decouple autocast and context-manager handling from Dynamo backends. Review notes that two of the APIs are ineffective and that the tests remain incomplete. Maintainers of external Dynamo backends may watch the change for a cleaner extension surface.

CUDA graph archive save proposed

A pull request opened in pytorch/pytorch proposes saving a captured CUDA graph to an archive. A bot comment marked the start of the work under number 199078. The feature would let users persist graph captures for later reload in optimized CUDA workflows.