●Change8

b8875

📦 llama-cppView on GitHub →
✨ 5 features🐛 4 fixes🔧 25 symbols

Summary

This release introduces comprehensive support for Reka Edge 2603, including a new Yasa2 vision encoder, a dedicated chat template handler, and GGUF conversion scripts. Numerous internal cleanups were performed, particularly around tensor handling and GGML operations.

Migration Steps

  1. When using the chat parser, use autoparser::generation_params instead of templates_params.
  2. Update image preprocessing settings for Reka Edge models to use image_resize_algo = RESIZE_ALGO_BICUBIC and image_resize_pad = false.

✨ New Features

  • Added support for Reka Edge 2603.
  • Added Yasa2 vision encoder support for mtmd (Reka Edge).
  • Added reka-edge template handler for chat (tools, thinking) using a PEG-based parser.
  • Added script to convert Reka Yasa2 HF checkpoints to GGUF format, including text decoder, mmproj (ConvNeXt + projection), and 2D sincos positional embeddings.
  • Added Reka Edge mixed quantization helper script for Q4_0 base quantization with Q8_0 override for the last 8 transformer blocks.

🐛 Bug Fixes

  • Fixed stray text appended in llama_decode_text by removing accidental concatenation of the full text string when formatting UNK_BYTE hex escapes.
  • Avoided duplicate tensor loading in the yasa2 vision encoder where TN_YASA_PATCH_W and TN_PATCH_EMBD resolved to the same tensor name, preventing memory pool overflow.
  • Simplified reasoning parser in chat-reka to match LFM2 pattern.
  • Removed image/video oracle tests due to incompatibility with oaicompat parser.

Affected Symbols

llama_decode_textPROJECTOR_TYPE_YASA2clip.cppmtmd.cppchat-reka.cppchat-reka.hReka-Edge.jinjatry_specialized_template()LLAMA_EXAMPLE_MTMDchat-template-fileautoparser::generation_paramstemplates_paramsTN_YASA_PATCH_WTN_PATCH_EMBDv.patch_embd.weightyasa2.cppggml_contggml_reshape_3dggml_reshape_4dggml_meanggml_addggml_mulggml_subggml_divggml_repeat