ggml-org/llama.cpp build b11970 is a new build of the C/C++-based LLM inference project. This release adds a convolution operation to the Hexagon backend. According to the release summary, double buffering and the data transfer path were improved as well. The release also includes a change that prevents overflow in half-precision storage and supports depthwise convolution. The source of this information is the b11970 release record in ggml-org/llama.cpp. The core of this update, therefore, is operation support in the Hexagon backend along with the related buffer, transfer, and storage changes.
The changes divide into an added operation and improvements to existing paths. The convolution operation was added to the Hexagon backend, and support for depthwise convolution is explicitly stated in the release summary. Separately, improvements to double buffering and the data transfer path, and the prevention of overflow in half-precision storage, are recorded. What can be confirmed from the summary is which items are reflected in b11970. Which models or usage conditions each item targets, and how behavior differs before and after the change, are not described in the provided information.
The release summary presents no performance figures, comparison results, or target models. Therefore, it cannot be judged from this information alone how convolution support or the buffer and transfer path improvements affect inference speed or memory usage. The overflow prevention for half-precision storage and the depthwise convolution support likewise come with no detailed conditions. The confirmable scope extends only to the fact that these changes are included in b11970. Concrete effects and scope of application can only be judged with additional material.