b9453
📦 llama-cppView on GitHub →
✨ 5 features🐛 4 fixes🔧 3 symbols
Summary
This release introduces full implementation support for EXAONE 4.5 models, including GQA for MMproj and specific vision token handling. Several fixes were applied related to EXAONE model loading and conversion.
Migration Steps
- If using EXAONE 4.5, note that image boundaries are marked by <vision> and </vision> tags, unlike Qwen which uses <|vision_start|> and <|vision_end|>.
✨ New Features
- Add EXAONE 4.5 implementations.
- Add GQA for MMproj.
- EXAONE 4.5 uses <vision> and </vision> for image boundaries; Qwen keeps <|vision_start|> and <|vision_end|>.
- Route EXAONE 4.5 through the Qwen2.5-VL-style encode path (window attention pattern, optional mmproj input norm).
- Update exaone4_5 projector weights and convert_hf_to_gguf for mmproj export.
🐛 Bug Fixes
- Fix EXAONE after merge.
- Fix EXAONE 4.5 conversion.
- Fix unintended deletion.
- Load EXAONE4 nextn tensors correctly by aligning registration with EXAONE_MOE for NextN/MTP slots and avoiding skip-flag propagation on duplicated rope_freqs so model loading succeeds for EXAONE 4.5 GGUF.