[gentoo-commits] repo/proj/guru:master commit in: sci-misc/llama-cpp/

["Lucio Sauer" <watermanpaint@posteo.net>]

commit:     19e74ee6a141256e081b6f4d7749ece9af97b307
Author:     Sergey Alirzaev <l29ah <AT> riseup <DOT> net>
AuthorDate: Fri Mar  7 11:16:01 2025 +0000
Commit:     Lucio Sauer <watermanpaint <AT> posteo <DOT> net>
CommitDate: Fri Mar  7 11:16:18 2025 +0000
URL:        https://gitweb.gentoo.org/repo/proj/guru.git/commit/?id=19e74ee6

sci-misc/llama-cpp: add 0_pre4848

for QwQ-32B --jinja support

Signed-off-by: Sergey Alirzaev <l29ah <AT> riseup.net>

 sci-misc/llama-cpp/Manifest                   |  1 +
 sci-misc/llama-cpp/llama-cpp-0_pre4848.ebuild | 93 +++++++++++++++++++++++++++
 2 files changed, 94 insertions(+)

diff --git a/sci-misc/llama-cpp/Manifest b/sci-misc/llama-cpp/Manifest
index 3085f9c75..aa174fbbd 100644
--- a/sci-misc/llama-cpp/Manifest
+++ b/sci-misc/llama-cpp/Manifest
@@ -1,2 +1,3 @@
 DIST llama-cpp-0_pre4576.tar.gz 20506059 BLAKE2B 8f011811e4df1f8d0c26b19f96a709980e078dc7e769b33cbbb03a852a29b489f80c8a1e298fecea53997068f6b7897e4536ba5db289aa445a1a6f16f98adce3 SHA512 21150721524283454ab53e370fdaf4e766f89fbb8d4b43072b10657d8c8b686630616cddbae7954147a2ba0360ad20c4643761f3774481e13a7b180812935c4e
 DIST llama-cpp-0_pre4763.tar.gz 20737582 BLAKE2B f6cb6885465e144c19698ac65410f59a6cc2b78d511968bc26c521ba90be87d102eb413e3ef903da30dae3336780e80a4c20cbbea30cc67375f790567e0e6e7a SHA512 c48923286e717d734a3414ae12182c869dd0a99fde722b46d48822a9cbcc5fc16ec5ade4108bd463990a3c9880ea58b559ba0a6975d04c348b474893df566bc9
+DIST llama-cpp-0_pre4848.tar.gz 20799416 BLAKE2B 6731dd8ab01f66fca24ad385c2611bc4a11df8beda692e88d79e1d6ab931c908d5a5d304d9c423d43c09d89e80c0b8dd1d57be23cb1a3bb522dbeac112ded604 SHA512 4e08cd24a96ad7c96abdc834d4b5d2d74ce01dd8774d81b693c25b890a0982ca9135cfb743b02f886277d2d0fc92c4e4b330e9acf0977fe00b2f1d4df70243d3

diff --git a/sci-misc/llama-cpp/llama-cpp-0_pre4848.ebuild b/sci-misc/llama-cpp/llama-cpp-0_pre4848.ebuild
new file mode 100644
index 000000000..b4db64b49
--- /dev/null
+++ b/sci-misc/llama-cpp/llama-cpp-0_pre4848.ebuild
@@ -0,0 +1,93 @@
+# Copyright 2025 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+ROCM_VERSION="6.3"
+
+inherit cmake rocm
+
+if [[ "${PV}" != "9999" ]]; then
+	KEYWORDS="~amd64"
+	MY_PV="b${PV#0_pre}"
+	S="${WORKDIR}/llama.cpp-${MY_PV}"
+	SRC_URI="https://github.com/ggerganov/llama.cpp/archive/refs/tags/${MY_PV}.tar.gz -> ${P}.tar.gz"
+else
+	inherit git-r3
+	EGIT_REPO_URI="https://github.com/ggerganov/llama.cpp.git"
+fi
+
+DESCRIPTION="Port of Facebook's LLaMA model in C/C++"
+HOMEPAGE="https://github.com/ggerganov/llama.cpp"
+
+LICENSE="MIT"
+SLOT="0"
+CPU_FLAGS_X86=( avx avx2 f16c )
+IUSE="curl openblas blis hip"
+REQUIRED_USE="?? ( openblas blis )"
+
+AMDGPU_TARGETS_COMPAT=(
+	gfx900
+	gfx90c
+	gfx902
+	gfx1010
+	gfx1011
+	gfx1012
+	gfx1030
+	gfx1031
+	gfx1032
+	gfx1034
+	gfx1035
+	gfx1036
+	gfx1100
+	gfx1101
+	gfx1102
+	gfx1103
+	gfx1150
+	gfx1151
+)
+
+# curl is needed for pulling models from huggingface
+# numpy is used by convert_hf_to_gguf.py
+DEPEND="
+	curl? ( net-misc/curl:= )
+	openblas? ( sci-libs/openblas:= )
+	blis? ( sci-libs/blis:= )
+	hip? (  >=dev-util/hip-6.3:= )
+"
+RDEPEND="${DEPEND}
+	dev-python/numpy
+"
+PATCHES=( "${FILESDIR}/blas-ld.diff" )
+
+src_configure() {
+	local mycmakeargs=(
+		-DLLAMA_BUILD_TESTS=OFF
+		-DLLAMA_BUILD_SERVER=ON
+		-DCMAKE_SKIP_BUILD_RPATH=ON
+		-DGGML_NATIVE=0	# don't set march
+		-DLLAMA_CURL=$(usex curl ON OFF)
+		-DBUILD_NUMBER="1"
+	)
+
+	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 hip; then
+		rocm_use_hipcc
+		mycmakeargs+=(
+			-DGGML_HIP=ON -DAMDGPU_TARGETS=$(get_amdgpu_flags)
+		)
+	fi
+
+	cmake_src_configure
+}