ggml-org/llama.cpp b11913 includes a change that makes it possible to specify the VTCM size for the Hexagon execution path directly and to append K·M·G suffixes to setting values. This is published on the GitHub release page. The repository description introduces llama.cpp as a C/C++-based LLM inference project. The change that can be confirmed in this release concerns the way the size is set for a specific execution path and how the input value is written. The source and version of the change are confirmed as the GitHub release page and b11913, respectively. From this information alone, there is no basis for saying that other features or behavior changed alongside it.
The core of this change is that the input determining the VTCM size in the Hexagon path can now be specified directly. The allowance of K·M·G suffixes on the value is also explicitly stated in the release summary. Therefore, what can be confirmed extends only to the size specification feature and the addition of suffix notation. The provided facts do not explain what VTCM refers to, in what format the setting value is entered, or what size each suffix corresponds to. The repository shows 132470 stars, but that figure alone cannot be used to judge the impact of this change or the suitability of the feature.
The release summary does not present the devices and usage conditions to which this setting applies, the multiplier basis of the suffixes, or the performance and memory impact of changing the size. The specific name of the setting item and the changed files are also not included among the confirmable facts. Therefore, the scope confirmed in b11913 is direct VTCM size specification in the Hexagon path and the allowance of K·M·G suffixes. Beyond that, the performance effects and usage conditions are difficult to judge from this release information alone.