mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2026-10-22 06:09:59 +00:00
* sd: sync with master-968-8cd731e * feat: add TAEQI2.1 support for Qwen Image 2.1 * feat: add MoSR upscaler support * feat: add Qwen Image 2.1 Turbo sigma preset * fix: reclaim before accepting wrapped Vulkan iGPU memory estimates * feat: add per-module automatic flash attention selection * feat: add automatic direct convolution selection * chore: move parameter and tokenizer diagnostics to debug * fix: prevent core dump in CPU backend if using FP8 (e4m3) tensors * fix: benchmark Vulkan direct convolution in auto mode * feat: add FLUX.3 Image Klein and Klein Base support * sd: update Makefile, fix flash attention workaround * sd: add support for Flux 3 * sd: sync with master-973-f0aac52 * fix: open model files with UTF-8 paths on Windows * fix: handle UTF-8 file paths on Windows * chore: sync ggml and remove obsolete conv1d workarounds * feat: add moss-image support * sd: add support for MOSS-Image * sd: sync with master-975-95b3090 * feat: support Diffusers-format MiniMax-H3 LoRAs * feat: support MOSS split QKV weights and batch host loading * sd: set conv_direct through the parameter list string |
||
|---|---|---|
| .. | ||
| bin | ||
| lib | ||
| packaging | ||
| expose.cpp | ||
| expose.h | ||
| gpttype_adapter.cpp | ||
| hash.cpp | ||
| json_to_gbnf.py | ||
| kcpp_backend.cpp | ||
| kcpp_backend.h | ||
| model_adapter.cpp | ||
| model_adapter.h | ||
| README.md | ||
| simplecpuinfo.cpp | ||
KoboldCpp native sources
This directory contains KoboldCpp-owned implementation and build-support code. Upstream-synced llama.cpp and ggml sources remain in their original directories so upstream updates can continue to be merged without path-only conflicts.
Prebuilt helper programs and link inputs used by KoboldCpp builds live under
bin/ and lib/, respectively. Release version metadata and its generation
scripts live under packaging/; Conda environment definitions are kept in
packaging/environments/.