May 2026 highlights for pytorch/pytorch: Implemented significant correctness and dtype-safety improvements in the Inductor convolution path, expanded edge-case handling, and extended dtype checks to in-place operations. The changes improve parity with eager semantics, reduce runtime errors, and enhance cross-device reliability, contributing to safer optimization and deployment workflows.

May 2025 monthly summary for jeejeelee/vllm. Focused on security hardening and stability: implemented FIPS-compliant MD5 hashing by applying usedforsecurity=False across the model loader, addressing a vulnerability in security-sensitive contexts. The change spans multiple classes, reflecting a comprehensive cross-module fix. This work reduces compliance risk and strengthens crypto-related operations in production.