The open‑source llama.cpp library released version b11505 on 2026‑10‑08. The update fixes a Vulkan shader bug in the top‑k operation. Previously, +inf and NaN values were ignored and could cause hangs on NVIDIA GPUs or wrong indices on AMD GPUs. The patch maps NaN to -inf, expands the bucket range to count every value, and corrects the k = 1 path for negative numbers. A new test called test_top_k_inf checks these cases.
Why it matters
Developers using llama.cpp with Vulkan will see fewer crashes and more accurate token‑ranking results when their prompts contain extreme float values.
