The OpenCL implementation has disabled the repacking of Flash Attention (FA) and Mixture of Experts (MoE) weights. This change is implemented to work around specific compiler issues affecting these operations.
OpenCL disables FA and MoE weights repack to workaround compiler issues
from English