# Copyright 1999-2026 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="LLM inference 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 -> ${P}.tar.gz
webui? (
https://github.com/ggml-org/llama.cpp/releases/download/${MY_PV}/llama-${MY_PV}-ui.tar.gz -> ${P}-ui.tar.gz
)
"
S="${WORKDIR}/llama.cpp-${MY_PV}"
KEYWORDS="~amd64 ~arm64"
fi
SRC_URI+="
examples? (
https://huggingface.co/ggml-org/tiny-llamas/resolve/${TINY_LLAMAS_COMMIT}/stories15M-q4_0.gguf
-> ggml-org_models_tinyllamas_stories15M-q4_0-${TINY_LLAMAS_COMMIT}.gguf
)
"
LICENSE="MIT"
SLOT="0"
X86_CPU_FLAGS=(
sse4_2
avx
avx_vnni
avx2
bmi2
avx512f avx512cd avx512vl avx512dq avx512bw
avx512vbmi
avx512_vnni
avx512_bf16
fma3
f16c
amx_tile
amx_int8
amx_bf16
)
CPU_FLAGS=( "${X86_CPU_FLAGS[@]/#/cpu_flags_x86_}" )
# ggml exposes no per-feature -DGGML_* option on ARM. The only lever is
# GGML_CPU_ARM_ARCH, an accumulated -march= string, so these flags are gated at
# compile time via -march tags rather than by individual cmake options.
# ggml SME (armv9.2-a) is deliberately not exposed: it has no cpu_flags_arm_*
# value in profiles/desc/cpu_flags_arm.desc, and forcing armv9.2-a would exclude
# every pre-Armv9 core. Users who want it set GGML_CPU_ARM_ARCH themselves.
ARM_CPU_FLAGS=( asimddp asimdhp sve i8mm sve2 )
CPU_FLAGS+=( "${ARM_CPU_FLAGS[@]/#/cpu_flags_arm_}" )
IUSE="openblas +openmp blis hip cuda opencl vulkan flexiblas wmma examples rpc +server webui ${CPU_FLAGS[*]}"
# The ROCm stack behind USE=hip (dev-util/hip, sci-libs/hipBLAS and
# sci-libs/rocWMMA) is ~amd64-only, so arm64? ( !hip ) makes the combination
# unselectable rather than leaving arm64 users with an unsatisfiable dependency.
#
# The flag is deliberately named "hip" and not "rocm". Empirically, naming it
# "rocm" makes pkgcheck report NonsolvableDeps{InDev,InStable} on arm64
# profiles while "hip" scans clean -- reproduced in both directions on this
# package and on sci-ml/whisper-cpp. The mechanism is NOT understood, and the
# pruning is not universal (an equivalent arm64? ( !webm ) guard in
# sci-ml/stable-diffusion-cpp does not silence libwebm). Treat the naming as a
# recorded empirical workaround, not a rule. See story 002 design D1.4a.
REQUIRED_USE="
?? ( openblas blis flexiblas )
hip? ( ${ROCM_REQUIRED_USE} )
wmma? ( hip )
webui? ( server )
arm64? ( !hip )
cpu_flags_arm_sve2? ( cpu_flags_arm_sve )
"
CDEPEND="
dev-libs/openssl
openmp? ( llvm-runtimes/openmp:= )
openblas? ( sci-libs/openblas:= )
blis? ( sci-libs/blis:= )
flexiblas? ( sci-libs/flexiblas:= )
hip? (
>=dev-util/hip-${ROCM_VERSION}
>=sci-libs/hipBLAS-${ROCM_VERSION}
wmma? ( >=sci-libs/rocWMMA-${ROCM_VERSION} )
)
cuda? ( dev-util/nvidia-cuda-toolkit:= )
"
DEPEND="${CDEPEND}
opencl? ( dev-util/opencl-headers )
vulkan? (
dev-util/spirv-headers
dev-util/vulkan-headers
)
"
RDEPEND="${CDEPEND}
opencl? ( dev-libs/opencl-icd-loader )
vulkan? ( media-libs/vulkan-loader )
"
BDEPEND="
vulkan? ( 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_unpack() {
if [[ ${PV} == *9999* ]]; then
git-r3_src_unpack
else
default
fi
if use webui; then
if [[ ${PV} == *9999* ]]; then
# Upstream publishes the prebuilt webui dist only per release
# tag. The mutable bucket pointer this used to fetch carries no
# Manifest checksum, so it cannot be expressed in SRC_URI and
# cannot be verified; fetching it here would also need
# RESTRICT="network-sandbox" for an unverifiable input. Every
# versioned snapshot carries the dist tarball as a real SRC_URI
# entry, so use one of those instead.
die "USE=webui is unsupported on a live ebuild: the webui dist has no verifiable source"
else
ln -s "${WORKDIR}/llama-${MY_PV}" "${S}/tools/ui/dist" || die
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() {
if [[ ${PV} == *9999* ]]; then
local mycmakeargs=(
-DLLAMA_BUILD_NUMBER="$(git rev-list --count HEAD)"
-DLLAMA_BUILD_COMMIT="$(git rev-parse HEAD)"
)
else
local mycmakeargs=( -DLLAMA_BUILD_NUMBER="${MY_PV#b}" )
fi
mycmakeargs+=(
-DGGML_CCACHE=OFF
-DCMAKE_SKIP_BUILD_RPATH=ON
-DLLAMA_BUILD_TESTS=OFF
-DLLAMA_BUILD_EXAMPLES="$(usex examples)"
-DLLAMA_BUILD_SERVER="$(usex server)"
-DLLAMA_BUILD_UI="$(usex webui)"
-DGGML_RPC="$(usex rpc)"
-DGGML_CUDA="$(usex cuda)"
-DGGML_OPENCL="$(usex opencl)"
-DGGML_OPENMP="$(usex openmp)"
-DGGML_VULKAN="$(usex vulkan)"
-DGGML_NATIVE=OFF
-DGGML_SSE42="$(usex cpu_flags_x86_sse4_2)"
-DGGML_AVX="$(usex cpu_flags_x86_avx)"
-DGGML_AVX_VNNI="$(usex cpu_flags_x86_avx_vnni)"
-DGGML_AVX2="$(usex cpu_flags_x86_avx2)"
-DGGML_BMI2="$(usex cpu_flags_x86_bmi2)"
-DGGML_AVX512_VBMI="$(usex cpu_flags_x86_avx512vbmi)"
-DGGML_AVX512_VNNI="$(usex cpu_flags_x86_avx512_vnni)"
-DGGML_AVX512_BF16="$(usex cpu_flags_x86_avx512_bf16)"
-DGGML_FMA="$(usex cpu_flags_x86_fma3)"
-DGGML_F16C="$(usex cpu_flags_x86_f16c)"
-DGGML_AMX_TILE="$(usex cpu_flags_x86_amx_tile)"
-DGGML_AMX_INT8="$(usex cpu_flags_x86_amx_int8)"
-DGGML_AMX_BF16="$(usex cpu_flags_x86_amx_bf16)"
)
if use cpu_flags_x86_avx512f &&
use cpu_flags_x86_avx512cd &&
use cpu_flags_x86_avx512vl &&
use cpu_flags_x86_avx512dq &&
use cpu_flags_x86_avx512bw; then
mycmakeargs+=( -DGGML_AVX512=ON )
else
mycmakeargs+=( -DGGML_AVX512=OFF )
fi
if use arm64; then
# Mirrors upstream's escalation table in ggml/src/ggml-cpu/CMakeLists.txt:
# each feature bumps the baseline to the first architecture level that
# supported it, and the tags accumulate. Order is load-bearing —
# last write wins, so the highest required level ends up in arm_arch.
# Setting GGML_CPU_ARM_ARCH bypasses that table entirely (it is the
# first branch of an if/elseif), which is why it is replicated here.
local arm_arch="armv8-a" arm_tags=""
if use cpu_flags_arm_asimddp; then arm_arch="armv8.2-a"; arm_tags+="+dotprod"; fi
if use cpu_flags_arm_asimdhp; then arm_arch="armv8.2-a"; arm_tags+="+fp16"; fi
if use cpu_flags_arm_sve; then arm_arch="armv8.2-a"; arm_tags+="+sve"; fi
if use cpu_flags_arm_i8mm; then arm_arch="armv8.6-a"; arm_tags+="+i8mm"; fi
if use cpu_flags_arm_sve2; then arm_arch="armv8.6-a"; arm_tags+="+sve2"; fi
mycmakeargs+=( -DGGML_CPU_ARM_ARCH="${arm_arch}${arm_tags}" )
fi
if use openblas; then
mycmakeargs+=(
-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
)
fi
if use blis; then
mycmakeargs+=(
-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FLAME
)
fi
if use flexiblas; then
mycmakeargs+=(
-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FlexiBLAS
)
fi
if use cuda; then
local -x CUDAHOSTCXX="$(cuda_gccdir)"
# tries to recreate dev symlinks
cuda_add_sandbox
addpredict "/dev/char/"
fi
if use hip; then
export HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)"
mycmakeargs+=(
-DAMDGPU_TARGETS="$(get_amdgpu_flags)"
-DGGML_HIP=ON
-DGGML_HIP_ROCWMMA_FATTN="$(usex wmma)"
)
fi
cmake_src_configure
}