v0.3.0
📦 llama-cppView on GitHub →
✨ 7 features🐛 9 fixes🔧 28 symbols
Summary
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.
Migration Steps
- The `fit` function has been moved out of the server and now takes `n_streams` into account.
- Removed the `-no-cnv` CLI option.
✨ New Features
- Introduced the dots3-note multimodal model with a new DSA-ISWA KV cache.
- Added MTP support for GLM-4.5-Air.
- Implemented tensor-split (`-sm tensor`) for DeepSeek 4.
- Added support for dots3-note vision and audio.
- Added WebP image decoding via ffmpeg.
- The server now includes a `LLAMA_SERVER_SLOTS_N_DIFF` debug knob.
- The web UI now features tabbed chat navigation.
🐛 Bug Fixes
- Fixed multi-sequence rollback issues for DeepSeek 4.
- Fixed meta tensor split state propagation for tensor parallel.
- Fixed DSpark support for bailingmoe3.
- Flattened in/out projections for mamba2 to dispatch GEMM instead of GEMV.
- Used `ggml_rope_set_offset` in deepseek2/4, dflash, minicpm3 and plm.
- Fixed parsing of `-` in grammar char classes to be a literal hyphen.
- Fixed clang LTO issues.
- Fixed draft-mtp with embeddings.
- Fixed loading videos with ffmpeg.
Affected Symbols
DSA-ISWA KV cacheGLM-4.5-AirDeepSeek 4ggml v0.22.0meta-backend tensor splitper-op Metal kernelsparallel compilationnon-in-place `ggml_clamp`mtmddots3-noteWebP decodingPillow-accurate resizeLLAMA_SERVER_SLOTS_N_DIFFbailingmoe3DSparkmamba2ggml_rope_set_offsetdeepseek2dflashminicpm3plmjson.hclang LTOfit functiondraft-mtpembeddingsffmpegLLAMA_SERVER_SLOTS_N_DIFF