llama.cpp v0.3.0 adds a new KV cache and MTP support

llama.cpp shipped version 0.3.0 on August 25, adding a DSA-ISWA KV cache type for the dots3-note model and multi-token prediction (MTP) support for GLM-4.5-Air, according to the project's v0.3.0 release notes. MTP lets GLM-4.5-Air generate more than one token per forward pass instead of one, speeding up local generation.

The update also adds tensor-split support for DeepSeek 4 through a new -sm tensor flag. It targets anyone spreading a model across multiple GPUs. A separate fix addresses multi-sequence generation, which could previously roll back to the wrong state.

Underneath it, ggml moves to v0.22.0, adding meta-backend tensor split and per-op Metal kernels that now compile in parallel. mtmd now decodes WebP images and resizes them to match Pillow's output pixel for pixel.

This is the project's third release on its semver track, tagged alongside the nightly bNNNN builds it has always cut.