The ggml-org team released llama.cpp version b12074 with changes to delta-net convolution state. The update makes the rollback slot index an I32 input filled on the host, rather than building it on the device with arange and argsort operations.
The release notes say the index is constant for each graph shape and that the change removes the need for the metadata backend to mirror a split for arange. They do not report a measured speed or memory improvement.
Why it matters
People running llama.cpp models that use delta-net may encounter a changed internal calculation, but the release notes do not establish whether it improves performance in practice.
