llama.cpp ebuild with rocm support

Signed-off-by: Igor Ulyanov <disinbox@gmail.com>

2 files changed, 79 insertions, 0 deletions

diff --git a/dev-cpp/llama-cpp/Manifest b/dev-cpp/llama-cpp/Manifest
new file mode 100644
index 0000000..8062f78
--- /dev/null
+++ b/dev-cpp/llama-cpp/Manifest
@@ -0,0 +1 @@
+EBUILD llama-cpp-9999.ebuild 1728 BLAKE2B 04f41b05672133b14185d4426d66a79a4f06309f0071c38e5767eb9f78ee81f567ef813804839bd987a2700bf3d48a3bb35ab8999923fa2c1dc30246252230ce SHA512 29482a3d961838dc525cd142699bc34357c8db4d17cb6ac9f99f86deb7ba5c3ab68fa302e148e943bd759a62189e08622ab77a6f62e53c7e35ac0ca1d0ef70fd
diff --git a/dev-cpp/llama-cpp/llama-cpp-9999.ebuild b/dev-cpp/llama-cpp/llama-cpp-9999.ebuild
new file mode 100644
index 0000000..ef4e288
--- /dev/null
+++ b/dev-cpp/llama-cpp/llama-cpp-9999.ebuild
@@ -0,0 +1,78 @@
+# Copyright 2023 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+ROCM_VERSION=5.5
+
+inherit cmake llvm rocm
+
+LLVM_MAX_SLOT=16
+
+EGIT_REPO_URI="https://github.com/ggerganov/llama.cpp.git"
+inherit git-r3
+
+DESCRIPTION="Port of Facebook's LLaMA model in C/C++"
+HOMEPAGE="https://github.com/ggerganov/llama.cpp"
+
+LICENSE="MIT"
+SLOT="0"
+KEYWORDS="~amd64"
+IUSE="blas cublas lto tests tools rocm"
+CPU_FLAGS_X86=( avx avx2 f16c )
+
+DEPEND="blas? ( sci-libs/openblas:= )
+	cublas? ( dev-util/nvidia-cuda-toolkit )
+	rocm? ( sci-libs/rocBLAS )"
+RDEPEND="${DEPEND}"
+BDEPEND="${DEPEND}"
+
+S="${WORKDIR}/${P}"
+
+src_configure() {
+	if use rocm ; then
+		CC=/usr/lib/llvm/${LLVM_MAX_SLOT}/bin/clang
+		CXX=/usr/lib/llvm/${LLVM_MAX_SLOT}/bin/clang++
+		export DEVICE_LIB_PATH=/usr/lib/amdgcn/bitcode
+		export HIP_DEVICE_LIB_PATH=/usr/lib/amdgcn/bitcode
+	fi
+	local mycmakeargs=(
+		-DLLAMA_BLAS="$(usex blas)"
+		-DLLAMA_CUBLAS="$(usex cublas)"
+		-DLLAMA_LTO="$(usex lto)"
+		-DLLAMA_BUILD_TESTS="$(usex tests)"
+		-DLLAMA_HIPBLAS="$(usex rocm)"
+		-DAMDGPU_TARGETS="$(get_amdgpu_flags)"
+		-DLLAMA_BUILD_SERVER=OFF
+		-DCMAKE_SKIP_BUILD_RPATH=ON
+		-DBUILD_NUMBER="1"
+	)
+	if use cublas ; then
+		addpredict /dev/nvidiactl
+	fi
+	cmake_src_configure
+}
+
+src_install() {
+	doheader llama.h
+
+	cd "${BUILD_DIR}" || die
+
+	dolib.so libllama.so
+
+	newbin bin/main llama-cpp
+
+	if use tools ; then
+		newbin bin/benchmark llama-cpp-benchmark
+		newbin bin/perplexity llama-cpp-perplexity
+		newbin bin/q8dot llama-cpp-q8dot
+		newbin bin/quantize llama-cpp-quantize
+		newbin bin/quantize-stats llama-cpp-quantize-stats
+		newbin bin/vdot llama-cpp-vdot
+	fi
+
+}
+
+pkg_postinst() {
+	elog "The main binary has been installed as \"llama-cpp\""
+}