The open‑source llama.cpp library released version b11632 on Oct 15 2026. It adds a chunked GDN (Gaussian‑Diffusion‑Normalization) kernel for both CUDA and HIP, improving GPU support. The update enables SM86 and SM89 instruction sets on tested Ampere and Ada GPUs, including RTX 3090 and RTX 4090, after correctness and prefill benchmarks. Older builds received fixes, and the code now uses a refactored swizzle API for the GDN MMA (matrix‑multiply‑accumulate) phase.
Why it matters
Developers can get faster, more reliable llama.cpp inference on recent Nvidia GPUs.
