# llama.cpp adds Vulkan KV dequantization optimization

> The open‑source Llama.cpp library now supports a dequantized key‑value cache on Vulkan GPUs, boosting speed for certain model heads.

Oossa · 2026-11-11 · https://oossa.com/en/llama-cpp-adds-vulkan-kv-dequantization-optimization

The ggml‑org team released llama.cpp version b12291. It adds a Vulkan‑based optimization that dequantizes the key‑value cache using the coopmat2 instruction. The change is applied only when a heuristic based on query tokens and GQA ratio deems it safe. The update also fixes alignment handling for some head sizes, improving performance when the optimization cannot be used.

## The facts

- Release tag b12291 published on 2026-11-11
- Adds dequantized KV optimization for Vulkan GPUs

## Why it matters

Developers running Llama models on Vulkan‑compatible hardware can see faster inference without changing their code.

## Sources & references

1. [ggml-org/llama.cpp b12291](https://github.com/ggml-org/llama.cpp/releases/tag/b12291) – llama.cpp, 2026-11-11

Last updated: 2026-11-11