sci-misc/koboldcpp
All-in-one local AI server (LLM, image, speech, TTS) on ggml/llama.cpp
Screenshots
View
Download
BrowseChangeLog
commit 7f5741cf5ac46ac8a8b76539f7711aef7d274b21
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 4 12:59:33 2026 +0200
sci-misc/koboldcpp: new package, add 1.116.1
All-in-one local AI server (LLM + Stable Diffusion + Whisper + TTS) built on
ggml/llama.cpp. Builds the ggml backends from the vendored source and installs
the pure-stdlib launcher (koboldcpp.py) with the embedded KoboldAI Lite web UI
and resources. The upstream PyInstaller blob and the bundled prebuilt glslc are
not used (Vulkan shaders build with the system media-libs/shaderc), and no
model weights are shipped.
USE=vulkan is upstream's official GPU path for both AMD and NVIDIA, and is
build-verified here (CPU + Vulkan). CUDA and ROCm are intentionally omitted for
now: CUDA cannot be verified on this host, and koboldcpp's ROCm support is an
unofficial upstream fork whose Makefile GPU_TARGETS handling needs work. Both
are follow-ups.
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 4 12:59:33 2026 +0200
sci-misc/koboldcpp: new package, add 1.116.1
All-in-one local AI server (LLM + Stable Diffusion + Whisper + TTS) built on
ggml/llama.cpp. Builds the ggml backends from the vendored source and installs
the pure-stdlib launcher (koboldcpp.py) with the embedded KoboldAI Lite web UI
and resources. The upstream PyInstaller blob and the bundled prebuilt glslc are
not used (Vulkan shaders build with the system media-libs/shaderc), and no
model weights are shipped.
USE=vulkan is upstream's official GPU path for both AMD and NVIDIA, and is
build-verified here (CPU + Vulkan). CUDA and ROCm are intentionally omitted for
now: CUDA cannot be verified on this host, and koboldcpp's ROCm support is an
unofficial upstream fork whose Makefile GPU_TARGETS handling needs work. Both
are follow-ups.