ggerganov/llama.cpp Released Version b11639
Release b11639 of llama.cpp introduces a targeted fix for OpenCL clamping on non-contiguous tensor views alongside standard compiled binaries. The update implements a new strided kernel to prevent incorrect element reads when processing non-contiguous input. This bug fix improves stability and correctness for users running large language models via OpenCL-supported hardware accelerators. Ensuring accurate tensor operations on non-contiguous views prevents subtle calculation errors during model inference. The update resolves an issue where the clamp kernel incorrectly indexed sources and destinations as flat contiguous arrays. A dedicated strided kernel, kernel_clamp_nc, is now invoked whenever the input source lacks contiguity.
## BACKGROUND
llama.cpp is an open-source software library used to perform inference on various large language models like Meta's Llama. OpenCL is an open, royalty-free standard for cross-platform parallel programming that allows developers to run intensive tasks across diverse hardware accelerators.