Build b12063 of ggml-org/llama.cpp has been released. According to the GitHub release summary, this build includes a change that moves dflash's hidden features to device copy. The source of this information is the b12063 release page of ggml-org/llama.cpp. So the core of this news is not the appearance of a new repository or an introduction of features in general, but this data-movement change recorded in that build. No grounds were given for adding anything about how the change is implemented or the conditions under which it applies beyond the scope stated in the release summary.
ggml-org/llama.cpp is a project that handles LLM inference in C/C++. This change was also introduced as inference-related work in this repository, but the provided description is limited to moving dflash's hidden features to device copy. Here, 'hidden features' is the expression used in the summary, and which value or data structure it refers to is not explained separately. Nor can it be determined from the given information alone between which devices the device copy is performed, or on which path it is used. Therefore the only confirmable change is the movement method stated in the release summary, and its background or specific scope should not be interpreted more broadly.
The provided release information contains no performance figures or comparison results for this change. Therefore it cannot be judged from this record alone what impact the change of moving to device copy has on inference performance. The meaning of dflash and the conditions under which the change applies were also not given. What can be confirmed in this release is the description of the change in b12063, and it is more appropriate to withhold conclusions about the actual effects or scope of application until separate grounds are available.