# llama.cpp adds non‑contiguous SUM support across all backends

> The ggml‑org/llama.cpp library now handles SUM operations on non‑contiguous tensors for CPU, CUDA, Vulkan and WebGPU.

Oossa · 2026-10-18 · https://oossa.com/en/llama-cpp-adds-non-contiguous-sum-support-across-all-backends

The open‑source llama.cpp library released a new version on 2026‑10‑18 that adds non‑contiguous SUM support. The change works on CPU, CUDA, Vulkan and WebGPU back‑ends, and the test suite now includes cases for these views. The update also expands the list of pre‑built binaries for macOS, Linux, Android and Windows, covering many hardware combos such as Apple Silicon, Intel, AMD, NVIDIA CUDA 12/13 and Snapdragon NPU. The project’s website and attestations are linked for download.

## The facts

- Release date: 2026‑10‑18 ("Sun Oct 18 2026 19:36:42 GMT+0200")
- New feature: non‑contiguous SUM support for CPU, CUDA, Vulkan, WebGPU

## Why it matters

Developers can now run faster, memory‑efficient summations on a wider range of devices, improving performance of LLM inference pipelines.

## Sources & references

1. [ggml-org/llama.cpp b11720](https://github.com/ggml-org/llama.cpp/releases/tag/b11720) – llama.cpp, 2026-10-18

Last updated: 2026-10-18