ggml-org/llama.cpp is an LLM inference project implemented in C/C++. The b11723 release, published on October 19, 2026, improves FLASH_ATTN_EXT in the Hexagon backend. The target of the change is the handling of long-context and sliding-window attention models. According to the GitHub release notes, the criteria for skipping masked blocks and switching compute paths were adjusted.
The purpose of this change is to reduce unnecessary data transfer in that model handling and to refine the selection of a compute path suited to the situation. Skipping blocks that a mask has been applied to translates into a reduction in transfer volume, and adjusting the switch criteria also changes how the compute path is chosen. What can therefore be confirmed in b11723 is an implementation change that touches a specific attention path in the Hexagon backend, and that is the scope of this release. There is no basis for describing it as a change to the whole project or to every backend.
However, the release notes do not separately verify any actual performance gain. How much processing speed or latency changed is not given in the provided information, so a reduction in data transfer alone is not enough to conclude that perceived performance improved. The verifiable scope covers the handling improvements to the Hexagon path for long-context and sliding-window attention models, and the content of that implementation. The performance effect remains unconfirmed until separate verification.