ggml-org/llama.cpp is a C/C++-based LLM inference project. In b12074, published on 2026-11-03, the recurrent memory convolution state was refactored into a reference, and the rewind slot number is now computed on the host instead of on the device. The source of the change is the b12074 release page of ggml-org/llama.cpp. What can be confirmed in this release is a change in how the state is handled and a change in where the slot number is computed. The two items are listed separately in the release summary. The summary does not describe them as new model support or the addition of user-facing features. Therefore the core of b12074 can be read as limited to changes to the recurrent memory state and the rewind computation path.
The recurrent memory convolution state has been refactored into a reference, and the rewind slot number computation has been moved from the device to the host. This simply points to the measures themselves as stated in the release summary. However, the summary alone does not reveal what concrete difference the reference-based approach makes internally, nor which device or execution path the slot number computation applies to. It also does not state what problem the change is meant to solve. It is more accurate not to assert the motivation or impact of the two measures beyond this scope.
The provided release information contains no performance figures and no before/after comparison. Changes in memory usage, latency, and stability are not presented either. There is therefore no basis for concluding that inference speed or stability improved because of this change alone. b12074 is a release that refactors the recurrent memory convolution state into a reference and changes where the rewind slot number is computed, and its effect is difficult to judge from the provided information alone.