b8035
A dangerous memory bug just got patched in the most popular LLM framework...
Deep Dive
The llama.cpp team released version b8035, a critical patch fixing a wrong memcpy length bug for block_interleave == 4. The update, signed with GitHub's verified signature, includes pre-built binaries for 22 different platforms including macOS (Apple Silicon/Intel), Windows (CUDA 12/13, Vulkan, SYCL, HIP), Linux (CPU/Vulkan), iOS, and openEuler. This addresses a stability and security vulnerability in the core inference library used by millions of developers and researchers worldwide.
Why It Matters
This patch prevents potential crashes and security exploits for anyone running local LLMs, making on-device AI more stable and secure.