# Copyright 1999-2026 Gentoo Authors
# Distributed under the terms of the GNU General Public License v2

EAPI=8

# Selects rocm.eclass's 10.x target list, the first with gfx1152/gfx1153.
# The hip/hipBLAS floor stays at 7.0 below: ROCM_USEDEP already requires a
# hipBLAS built for every enabled target.
ROCM_VERSION="10.0"
PYTHON_COMPAT=( python3_{12..15} )

inherit cmake cuda flag-o-matic git-r3 rocm linux-info python-any-r1 toolchain-funcs

TINY_LLAMAS_COMMIT="99dd1a73db5a37100bd4ae633f4cfce6560e1567"

DESCRIPTION="Port of Facebook's LLaMA model in C/C++"
HOMEPAGE="https://github.com/ggml-org/llama.cpp"

EGIT_REPO_URI="https://github.com/ggml-org/llama.cpp.git"

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=(
	amx_bf16 amx_int8 amx_tile avx avx2 avx512_bf16 avx512_vnni avx512f
	avx512vbmi avx_vnni bmi2 f16c fma3 sse4_2
)

IUSE="debug openblas +openmp blis rocm cuda opencl +openssl vulkan flexiblas examples +webui"
IUSE+=" ${CPU_FLAGS_X86[@]/#/cpu_flags_x86_}"

# HIP builds the CUDA backend sources, and ggml registers that backend once.
REQUIRED_USE="
	?? (
		openblas
		blis
		flexiblas
	)
	?? ( cuda rocm )
	rocm? ( ${ROCM_REQUIRED_USE} )
"

# No pinned UI distfile exists for a live checkout; CMake fetches it from
# Hugging Face at build time.
RESTRICT="webui? ( network-sandbox )"

# No sycl flag: the SYCL backend needs Intel's icpx, MKL's SYCL API and a
# GPU-enabled oneDNN, none of which ::gentoo packages. checked 2026-10-19
CDEPEND="
	openblas? ( sci-libs/openblas:= )
	openmp? ( llvm-runtimes/openmp:= )
	blis? ( sci-libs/blis:= )
	flexiblas? ( sci-libs/flexiblas:= )
	rocm? (
		>=dev-util/hip-7.0:=
		>=sci-libs/hipBLAS-7.0:=[${ROCM_USEDEP}]
		>=sci-libs/rocBLAS-7.0:=[${ROCM_USEDEP}]
	)
	cuda? ( dev-util/nvidia-cuda-toolkit:= )
	openssl? ( dev-libs/openssl:= )
	opencl? ( dev-libs/opencl-icd-loader )
	vulkan? ( media-libs/vulkan-loader )
"
DEPEND="${CDEPEND}
	opencl? ( dev-util/opencl-headers )
	vulkan? (
		dev-util/spirv-headers
		dev-util/vulkan-headers
	)
"
RDEPEND="${CDEPEND}"
# The OpenCL backend embeds its kernels with a stdlib-only Python script.
# The BLAS backend finds its library through pkg-config.
# The GPU backends compile with the build host's hipcc or nvcc.
BDEPEND="
	opencl? ( ${PYTHON_DEPS} )
	vulkan? ( media-libs/shaderc )
	openblas? ( virtual/pkgconfig )
	blis? ( virtual/pkgconfig )
	flexiblas? ( virtual/pkgconfig )
	rocm? ( >=dev-util/hipcc-7.0 )
	cuda? ( dev-util/nvidia-cuda-toolkit )
"

pkg_pretend() {
	[[ ${MERGE_TYPE} != binary ]] && use openmp && tc-check-openmp
}

pkg_setup() {
	[[ ${MERGE_TYPE} != binary ]] && use openmp && tc-check-openmp
	use opencl && python-any-r1_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
}

PATCHES=(
	# Stock LLVM 23+ rejects __bf16 vectors for the bf16 MFMA builtins;
	# upstream fixed the WMMA half in 9145ddb955. verified 2026-10-26
	"${FILESDIR}"/${PN}-0.7.0_p11906-hip-bf16-mfma.patch
)

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() {
	# Upstream releases build Release, which supplies -DNDEBUG; cmake.eclass
	# blanks CMAKE_*_FLAGS_RELWITHDEBINFO, so it never lands and the plain
	# assert() calls in ggml and llama stay live -- a configuration upstream
	# never exercises. GGML_ASSERT is unconditional, so the ggml consistency
	# checks survive either way. Worth ~1% of token throughput on AVX-512.
	# USE=debug keeps the asserts for diagnosing an abort. # verified 2026-09-21
	if ! use debug; then
		append-cflags "-DNDEBUG"
		append-cxxflags "-DNDEBUG"
	fi

	local mycmakeargs=(
		-DLLAMA_BUILD_TESTS=OFF
		-DLLAMA_BUILD_EXAMPLES=$(usex examples)
		-DLLAMA_BUILD_SERVER=ON
		-DCMAKE_SKIP_BUILD_RPATH=ON
		-DGGML_NATIVE=0	# don't set march
		-DGGML_RPC=ON
		# Otherwise RPC links libibverbs whenever the build host has it.
		-DGGML_RPC_RDMA=OFF
		# ggml would wrap the compiler in any ccache it finds; Portage's
		# FEATURES=ccache decides that instead.
		-DGGML_CCACHE=OFF
		-DLLAMA_OPENSSL=$(usex openssl)
		-DGENTOO_REMOVE_CMAKE_BLAS_HACK=ON
		-DGGML_CUDA=$(usex cuda)
		-DGGML_CUDA_NCCL=OFF
		-DGGML_OPENCL=$(usex opencl)
		# ggml accepts only Adreno and Intel GPUs, and its Adreno kernels reject
		# every other device, so they are enabled only where Adreno exists.
		# verified 2026-10-06
		-DGGML_OPENCL_USE_ADRENO_KERNELS=$(usex arm64)
		-DGGML_OPENMP=$(usex openmp)
		-DGGML_VULKAN=$(usex vulkan)
		-DGGML_SYCL=OFF

		# Isolate libraries shared with whisper.cpp.
		-DCMAKE_INSTALL_LIBDIR="${EPREFIX}/usr/$(get_libdir)/llama.cpp"
		-DCMAKE_INSTALL_RPATH="${EPREFIX}/usr/$(get_libdir)/llama.cpp"
	)

	# USE=webui fetches upstream's prebuilt UI from Hugging Face (the
	# checkout's build number, else "latest"). The npm build stays off:
	# it would run only when the build host happens to have npm.
	# A live checkout keeps git describe's version.
	mycmakeargs+=(
		-DLLAMA_BUILD_UI=OFF
		-DLLAMA_USE_PREBUILT_UI=$(usex webui)
	)

	mycmakeargs+=(
		-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=$(usex cpu_flags_x86_bmi2)
		-DGGML_F16C=$(usex cpu_flags_x86_f16c)
		-DGGML_FMA=$(usex cpu_flags_x86_fma3)
		-DGGML_AVX512=$(usex cpu_flags_x86_avx512f)
		-DGGML_AVX512_VBMI=$(usex cpu_flags_x86_avx512vbmi)
		-DGGML_AVX512_VNNI=$(usex cpu_flags_x86_avx512_vnni)
		-DGGML_AVX512_BF16=$(usex cpu_flags_x86_avx512_bf16)
		-DGGML_AMX_TILE=$(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 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)/g++"
		# nvcc recreates device symlinks during detection.
		cuda_add_sandbox
		addpredict "/dev/char/"
	fi

	if use rocm; then
		rocm_use_hipcc
		mycmakeargs+=(
			-DGGML_HIP=ON -DAMDGPU_TARGETS=$(get_amdgpu_flags)
		)
		# hipcc 10.1 hides libomp from CMake; see rocm_hipcc_openmp_args.
		use openmp && mycmakeargs+=( $(rocm_hipcc_openmp_args) )
	fi

	cmake_src_configure
}

src_compile() {
	cmake_src_compile

	# A failed fetch only warns and builds the server without a UI. Assets
	# placed in the checkout's tools/ui/dist (e.g. by a user patch) are
	# embedded from there instead.
	if use webui && [[ ! -f ${BUILD_DIR}/tools/ui/dist/index.html &&
		! -f ${S}/tools/ui/dist/index.html ]]; then
		die "USE=webui: the prebuilt UI could not be fetched from Hugging Face"
	fi
}

src_install() {
	cmake_src_install

	# Both projects install conflicting ggml headers.
	rm -r "${ED}/usr/include" || die
}