Release b11804 of ggml-org/llama.cpp adds a feature that extracts integer indices from a 32-bit float array in OpenCL operations. The repository is a C/C++-based LLM inference project. This change is a release that adds that operation on the OpenCL side, and the source is the b11804 release of ggml-org/llama.cpp. The changes listed in the release material are this feature and the initial value setting.
The reason the initial value was set to 0 in this change is to produce the same results as the CPU. So the newly added points are summarized as an OpenCL operation that obtains integer indices from a float array, and the handling of the initial value so that its results match the CPU. What can be confirmed here is the addition of the operation and the initial value setting; it is not described as the addition of any specific model or user-facing feature. The scope the repository covers is, by description, LLM inference in C/C++.
The material presents no information about the models or features that use this operation, the scope of application, or performance changes. Since no other changes are listed in the provided release summary either, there is no basis for interpreting it as a broader feature addition. Therefore this release alone cannot be used to judge the impact or speed difference in actual inference. The confirmable changes extend to the addition of the OpenCL operation and the initial value setting that matches the CPU results. The scope of impact can only be judged with further information.