sci-misc/llama-cpp: bump up to 6710, rm old

LFM2-MoE support and fixes Signed-off-by: Sergey Alirzaev <l29ah@riseup.net>
2026-07-18 03:23:20 -04:00 · 2025-10-08 04:10:18 +02:00
parent 14c9dde9e2
commit acfe4100ae
3 changed files with 126 additions and 94 deletions
--- a/sci-misc/llama-cpp/llama-cpp-0_pre6710.ebuild
+++ b/sci-misc/llama-cpp/llama-cpp-0_pre6710.ebuild
@@ -0,0 +1,125 @@
+# Copyright 2025 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
+
+if [[ "${PV}" != "9999" ]]; then
+	KEYWORDS="~amd64"
+	MY_PV="b${PV#0_pre}"
+	S="${WORKDIR}/llama.cpp-${MY_PV}"
+	SRC_URI="https://github.com/ggml-org/llama.cpp/archive/refs/tags/${MY_PV}.tar.gz -> ${P}.tar.gz"
+else
+	inherit git-r3
+	EGIT_REPO_URI="https://github.com/ggml-org/llama.cpp.git"
+fi
+
+DESCRIPTION="Port of Facebook's LLaMA model in C/C++"
+HOMEPAGE="https://github.com/ggml-org/llama.cpp"
+
+LICENSE="MIT"
+SLOT="0"
+CPU_FLAGS_X86=( avx avx2 f16c )
+IUSE="curl openblas +openmp blis hip cuda opencl vulkan"
+REQUIRED_USE="?? ( openblas blis )"
+
+# curl is needed for pulling models from huggingface
+# numpy is used by convert_hf_to_gguf.py
+CDEPEND="
+	curl? ( net-misc/curl:= )
+	openblas? ( sci-libs/openblas:= )
+	openmp? ( llvm-runtimes/openmp:= )
+	blis? ( sci-libs/blis:= )
+	hip? ( >=dev-util/hip-6.3:=
+		>=sci-libs/hipBLAS-6.3:=
+	)
+	cuda? ( dev-util/nvidia-cuda-toolkit:= )
+"
+DEPEND="${CDEPEND}
+	opencl? ( dev-util/opencl-headers )
+	vulkan? ( dev-util/vulkan-headers )
+"
+RDEPEND="${CDEPEND}
+	dev-python/numpy
+	opencl? ( dev-libs/opencl-icd-loader )
+	vulkan? ( media-libs/vulkan-loader )
+"
+BDEPEND="media-libs/shaderc"
+
+pkg_setup() {
+	if use hip; then
+		linux-info_pkg_setup
+		if linux-info_get_any_version && linux_config_exists; then
+			if ! linux_chkconfig_present HSA_AMD_SVM; then
+				ewarn "To use ROCm/HIP, you need to have HSA_AMD_SVM option enabled in your kernel."
+			fi
+		fi
+
+	fi
+}
+
+src_prepare() {
+	use cuda && cuda_src_prepare
+
+	cmake_src_prepare
+}
+
+src_configure() {
+	local mycmakeargs=(
+		-DLLAMA_BUILD_TESTS=OFF
+		-DLLAMA_BUILD_SERVER=ON
+		-DCMAKE_SKIP_BUILD_RPATH=ON
+		-DGGML_NATIVE=0	# don't set march
+		-DGGML_RPC=ON
+		-DLLAMA_CURL=$(usex curl ON OFF)
+		-DBUILD_NUMBER="1"
+		-DGENTOO_REMOVE_CMAKE_BLAS_HACK=ON
+		-DGGML_CUDA=$(usex cuda ON OFF)
+		-DGGML_OPENCL=$(usex opencl ON OFF)
+		-DGGML_OPENMP=$(usex openmp ON OFF)
+		-DGGML_VULKAN=$(usex vulkan ON OFF)
+
+		# 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 cuda; then
+		local -x CUDAHOSTCXX="$(cuda_gccdir)"
+		# tries to recreate dev symlinks
+		cuda_add_sandbox
+		addpredict "/dev/char/"
+	fi
+
+	if use hip; then
+		rocm_use_hipcc
+		mycmakeargs+=(
+			-DGGML_HIP=ON -DAMDGPU_TARGETS=$(get_amdgpu_flags)
+		)
+	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"
+}