PyTorch UAF report and mmap RFC

PyTorch developers reported a use-after-free issue and floated an RFC for optional memory-mapped checkpoint loading. Both items touch core runtime and storage paths that matter for model execution and distributed training.

Use-after-free in Dynamo guard invalidation

A report in the PyTorch repository describes a use-after-free in dynamo guard invalidation paths that rely on unlocked raw pointers. The issue sits in security-sensitive guard handling and can affect correctness or stability when guards are invalidated. Readers tracking PyTorch runtime safety should note the unlocked pointer pattern as a concrete failure mode in that subsystem.

RFC for opt-in mmap loading in FileSystemReader

PyTorch opened an RFC proposing opt-in mmap loading for the local FileSystemReader. The change aims to reduce eager materialization of tensor storage when reading .distcp files. The proposal targets lower memory pressure during checkpoint loads for users who choose the mmap path.