commit 2fa9afe33b44fe187fa95a8c276a6b89e4bb95e9
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Jun 18 07:35:32 2026 +0200

sci-ml/ollama: new package, add 0.30.10

Local runner for Llama, Mistral, Gemma and other LLMs. Forked from guru,
which lags at 0.21.2; 0.30.x reworked the build into a CMake superbuild that
compiles the GGML/llama.cpp inference backends as nested projects.

The llama.cpp commit pinned in the upstream LLAMA_CPP_VERSION file (b9672) is
prestaged as a distfile and wired in via FETCHCONTENT_SOURCE_DIR_LLAMA_CPP so
the build needs no network. The Ollama llama.cpp compat set (llama/compat) is
applied to that tree in src_prepare, including the models/ architecture
patches whose sources are linked into the llama target. The Go binary is built
by the superbuild against the gentoo-golang-dist dependency tarball.

USE cuda/rocm/vulkan select the cuda_v13 / rocm_v7_2 / vulkan llama-server
backends; the CPU backend builds all microarch variants and the right one is
loaded at runtime, so no cpu_flags_x86 flags are needed.

Two local patches drop upstream's bundling of the CUDA/ROCm runtime libraries
into the payload (Gentoo provides them via nvidia-cuda-toolkit / rocBLAS +
hipBLAS on the loader path) and remove a hipcc-only -parallel-jobs flag that
the upstream LLVM clang used as the HIP compiler rejects.