commit: fe3e646a7e00b5995b8e0d5d03dea436ec75b7d8 Author: Paul Zander <negril.nx+gentoo <AT> gmail <DOT> com> AuthorDate: Mon Jan 26 14:13:23 2026 +0000 Commit: Paul Zander <negril.nx+gentoo <AT> gmail <DOT> com> CommitDate: Mon Jan 26 14:13:23 2026 +0000 URL: https://gitweb.gentoo.org/repo/proj/guru.git/commit/?id=fe3e646a
sci-misc/llama-cpp: add 0_pre7836 Signed-off-by: Paul Zander <negril.nx+gentoo <AT> gmail.com> sci-misc/llama-cpp/Manifest | 1 + sci-misc/llama-cpp/llama-cpp-0_pre7836.ebuild | 162 ++++++++++++++++++++++++++ 2 files changed, 163 insertions(+) diff --git a/sci-misc/llama-cpp/Manifest b/sci-misc/llama-cpp/Manifest index 3fdd253743..5c86cf9090 100644 --- a/sci-misc/llama-cpp/Manifest +++ b/sci-misc/llama-cpp/Manifest @@ -5,3 +5,4 @@ DIST llama-cpp-0_pre6980.tar.gz 26431911 BLAKE2B b7d7c0dcdabde01acb816e73bc34456 DIST llama-cpp-0_pre7276.tar.gz 27765814 BLAKE2B d0553ab1dd29c9d93a18c6217aab4553faf09e385a94b90732a537bbcf9bded54d5cda28553543e2c0cc71b6a157bfb80a48405f3f8281c51525757967b33e16 SHA512 3035fe53fea2ca3b0f35e479f4eaec75e38a2ea670600445776cd6fa696fc83ca19eb6dd7cd2ab1da69e78293c62318b5182e6e5b3423ae6c1f00854c5132a4c DIST llama-cpp-0_pre7611.tar.gz 28622786 BLAKE2B 3c345645c9bcf07d8a513b9e883619b31b5254581f73429d638403758429fd2dfc5f78a22d538e8d88eb6c1be74bf805481af697480727ed750492ddec5c37fe SHA512 c6c4780d7e68adfc385b57c6f7530423f8205bfa283572b0d414d55e143c03307e98676e41ad527c37d7837f831f8ff24be0f7bf59e366ea82f3802cdc946821 DIST llama-cpp-0_pre7770.tar.gz 28797089 BLAKE2B 0ad614f16c19ff1339571dd90be566ff4ccedfd991dfeb948f0cedc54f8447a3e72e35ac392bb60d5bc8a44d1757be3bea0123eac02b694f5f0a8f2c2b941b6a SHA512 78896fdcf05330bb4b1fd86a985da56882d0166c6f276afda273dca183e4c7365decd5e9630c8d633e065162cbb06d6d8e4fdcf76be9768b238ee5053abaa3aa +DIST llama-cpp-0_pre7836.tar.gz 28813563 BLAKE2B c843c0199b528114c23c58536552e2a7f581fda93364353e8aa38fda99ddb1942c421ad00be4daf86b49a2c311ef89bb6908b8900ca1ffb64df338cbfa11c354 SHA512 161176107de175d9b6fdc1ebfa9fe0cf6fd6968245730f7df7f17c6a179ef111685d0fed5a45031825dc4b07db270887a4307061f9918e375a1c65f261062d9d diff --git a/sci-misc/llama-cpp/llama-cpp-0_pre7836.ebuild b/sci-misc/llama-cpp/llama-cpp-0_pre7836.ebuild new file mode 100644 index 0000000000..ff61000fd0 --- /dev/null +++ b/sci-misc/llama-cpp/llama-cpp-0_pre7836.ebuild @@ -0,0 +1,162 @@ +# Copyright 2026 Gentoo Authors +# Distributed under the terms of the GNU General Public License v2 + +EAPI=8 + +ROCM_VERSION="6.3" + +inherit cmake cuda rocm linux-info + +TINY_LLAMAS_COMMIT="99dd1a73db5a37100bd4ae633f4cfce6560e1567" + +DESCRIPTION="Port of Facebook's LLaMA model in C/C++" +HOMEPAGE="https://github.com/ggml-org/llama.cpp"; + +if [[ ${PV} == *9999* ]]; then + inherit git-r3 + EGIT_REPO_URI="https://github.com/ggml-org/llama.cpp.git"; +else + MY_PV="b${PV#0_pre}" + SRC_URI="https://github.com/ggml-org/llama.cpp/archive/refs/tags/${MY_PV}.tar.gz -> ${P}.tar.gz" + S="${WORKDIR}/llama.cpp-${MY_PV}" + KEYWORDS="~amd64" +fi + +SRC_URI+=" + examples? ( + https://huggingface.co/ggml-org/tiny-llamas/resolve/${TINY_LLAMAS_COMMIT}/stories15M-q4_0.gguf + -> ggml-org_models_tinyllamas_stories15M-q4_0-${TINY_LLAMAS_COMMIT}.gguf + ) +" + +LICENSE="MIT" +SLOT="0" +CPU_FLAGS_X86=( avx avx2 f16c ) + +# wwma USE explained here: https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md#hip +IUSE="curl openblas +openmp blis rocm cuda opencl vulkan flexiblas wmma examples" + +REQUIRED_USE=" + ?? ( + openblas + blis + flexiblas + ) + wmma? ( + rocm + ) +" + +# curl is needed for pulling models from huggingface +# numpy is used by convert_hf_to_gguf.py +CDEPEND=" + curl? ( net-misc/curl:= ) + openblas? ( sci-libs/openblas:= ) + openmp? ( llvm-runtimes/openmp:= ) + blis? ( sci-libs/blis:= ) + flexiblas? ( sci-libs/flexiblas:= ) + rocm? ( + >=dev-util/hip-${ROCM_VERSION}:= + >=sci-libs/hipBLAS-${ROCM_VERSION}:= + wmma? ( + >=sci-libs/rocWMMA-${ROCM_VERSION}:= + ) + ) + cuda? ( dev-util/nvidia-cuda-toolkit:= ) +" +DEPEND="${CDEPEND} + opencl? ( dev-util/opencl-headers ) + vulkan? ( dev-util/vulkan-headers ) +" +RDEPEND="${CDEPEND} + dev-python/numpy + opencl? ( dev-libs/opencl-icd-loader ) + vulkan? ( media-libs/vulkan-loader ) +" +BDEPEND="media-libs/shaderc" + +pkg_setup() { + if use rocm; then + linux-info_pkg_setup + if linux-info_get_any_version && linux_config_exists; then + if ! linux_chkconfig_present HSA_AMD_SVM; then + ewarn "To use ROCm/HIP, you need to have HSA_AMD_SVM option enabled in your kernel." + fi + fi + fi +} + +src_prepare() { + use cuda && cuda_src_prepare + cmake_src_prepare + if use examples; then + mkdir -p "${BUILD_DIR}/tinyllamas" || die + cp "${DISTDIR}/ggml-org_models_tinyllamas_stories15M-q4_0-${TINY_LLAMAS_COMMIT}.gguf" \ + "${BUILD_DIR}/tinyllamas/stories15M-q4_0.gguf" || die + fi +} + +src_configure() { + local mycmakeargs=( + -DLLAMA_BUILD_TESTS=OFF + -DLLAMA_BUILD_EXAMPLES=$(usex examples) + -DLLAMA_BUILD_SERVER=ON + -DCMAKE_SKIP_BUILD_RPATH=ON + -DGGML_NATIVE=0 # don't set march + -DGGML_RPC=ON + -DLLAMA_CURL=$(usex curl) + -DBUILD_NUMBER="1" + -DGENTOO_REMOVE_CMAKE_BLAS_HACK=ON + -DGGML_CUDA=$(usex cuda) + -DGGML_OPENCL=$(usex opencl) + -DGGML_OPENMP=$(usex openmp) + -DGGML_VULKAN=$(usex vulkan) + + # avoid clashing with whisper.cpp + -DCMAKE_INSTALL_LIBDIR="${EPREFIX}/usr/$(get_libdir)/llama.cpp" + -DCMAKE_INSTALL_RPATH="${EPREFIX}/usr/$(get_libdir)/llama.cpp" + ) + + if use openblas ; then + mycmakeargs+=( + -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS + ) + fi + + if use blis ; then + mycmakeargs+=( + -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FLAME + ) + fi + + if use flexiblas; then + mycmakeargs+=( + -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FlexiBLAS + ) + fi + + if use cuda; then + local -x CUDAHOSTCXX="$(cuda_gccdir)" + # tries to recreate dev symlinks + cuda_add_sandbox + addpredict "/dev/char/" + fi + + if use rocm; then + rocm_use_hipcc + mycmakeargs+=( + -DGGML_HIP=ON -DAMDGPU_TARGETS=$(get_amdgpu_flags) + -DGGML_HIP_ROCWMMA_FATTN=$(usex wmma) + ) + fi + + cmake_src_configure +} + +src_install() { + cmake_src_install + dobin "${BUILD_DIR}/bin/rpc-server" + + # avoid clashing with whisper.cpp + rm -rf "${ED}/usr/include" +}
