ggml-org/llama.cpp is an LLM inference project implemented in C/C++. Release b12065 fixes a bug where a shared key-value cache with no cells returned a row count of 0 during server warmup. This change is recorded in the b12065 release of ggml-org/llama.cpp. The core of this update is the correction of the behavior that returned an incorrect row count for the shared cache in the server warmup path. It is not a new feature or an added model, but a fix for the cache row count issue named in the release summary.
The impact of the issue did not end with the cache returning 0 rows. Because of this return value, processing could fall back to a dense attention mask, and after the scheduler replanned, it could end up being aborted under a larger mask. Therefore, the fix targets the row count returned by the shared key-value cache during server warmup, together with the mask selection and scheduling flow that this value feeds into. The release summary describes this sequence together, showing how a 0-row return can pass through a mask switch and replanning and end in an abort.
The confirmed scope is the row count handling issue of the shared key-value cache with no cells during server warmup. The provided release information presents no reproduction procedure, no performance figures, and no range of affected configurations. It is therefore appropriate to understand this fix as correcting the error in that warmup path, and there is no basis for extending the claim that the problem has been solved in other paths as well.