Build b11827 of ggml-org/llama.cpp has been published. The repository is an LLM inference project based on C/C++, and this release adds an OpenCL depthwise convolution for Adreno. The release summary also records changes to use a local program approach and to refine argument types. The core of this update therefore lies in the new convolution path, the program approach, and the argument types. The published notes give the version as b11827, and the source of the changes is the GitHub release page.
This change concerns the OpenCL path for Adreno within llama.cpp's inference implementation. The scope given by the repository description is LLM inference in C/C++, and this release summary explicitly states that a depthwise convolution was added to it. The notes also mention that the local program approach and argument types were refined, but they do not say which code or execution conditions each fix applies to. There is therefore no basis for asserting that this expands support on specific models or devices, or replaces existing behavior. The published facts are limited to the names and targets of the changed items.
The release information provides no performance measurements or comparison results, so the effect of the new path on inference speed cannot be judged. Which Adreno devices and models can use it, and what difference the change to the local program approach makes in actual execution, cannot be known from the summary alone. What this record confirms is only that related code changes went into b11827. It is therefore better to limit the claim to the changes themselves, rather than accept performance gains or the applicable scope as confirmed results, until concrete conditions and measurements are published.