dflash

llama.cpp 0.3.0 introduces multimodal capabilities with the dots3-note model and MTP support for GLM-4.5-Air, alongside significant updates to ggml and core model handling.

This release introduces support for Deepseek 4 with '-sm tensor' and includes various improvements such as coarser head split granularity and enhanced model saving capabilities. It also addresses a bug in dspark.

This release focuses on improving the clarity of logging for target_layer_ids in the dflash utility. The output format has been updated for better readability.

Enabled backend sampling for dflash and dspark, including adjustments to p_min in backend sampling.

This release introduces DFlash v2 support, including sliding window attention configuration, and provides updated pre-compiled binaries for numerous platforms and hardware accelerators.