bark-cpp.spec

# bark.cpp + encodec.cpp on the system ggml package (libggml +
# dlopen backends). Do not compile bundled ggml.

# Out-of-tree cmake/ninja can leave empty debugsourcefiles.list.
%undefine _debugsource_packages

%bcond_without examples
%if %{with examples}
%global build_examples ON
%else
%global build_examples OFF
%endif

# 1.0.0 + encodec sync (chore #199). No newer tag.
%global commit         5d5be84f089ab9ea53b7a793f088d3fbf7247495
%global encodec_commit 1cc279db4da979455651fbac1cbd151a2d121609

Summary:		Suno Bark text-to-speech in C/C++ (bark.cpp)
Name:			bark-cpp
Version:		1.0.0
Release:		4
License:		MIT
Group:			Sciences/Other
URL:			https://github.com/PABannier/bark.cpp
Source0:		https://github.com/PABannier/bark.cpp/archive/%{commit}/bark.cpp-%{commit}.tar.gz
# Vendored as encodec.cpp/ (git submodule). ggml/ is dropped — system ggml.
Source1:		https://github.com/PABannier/encodec.cpp/archive/%{encodec_commit}/encodec.cpp-%{encodec_commit}.tar.gz
Patch0:			0001-system-ggml.patch

# Prefer -O3 over distro -Os for the inference hot path
%global optflags %{optflags} -O3

BuildRequires:	cmake(ggml) >= 0.22.0
BuildRequires:	git-core

Requires:	%{mklibname ggml}%{?_isa} >= 0.22.0
Recommends:	ggml-backend-blas%{?_isa}
Recommends:	ggml-backend-vulkan%{?_isa}
Suggests:	ggml-backend-opencl%{?_isa}
Suggests:	ggml-backend-hip%{?_isa}

# ggml-config.cmake lists optional backends as hard deps
%global __requires_exclude cmake\\((hip|roc|mkl|intelsycl|cudatoolkit|CUDAToolkit|dnnl|DNNL|openvino|OpenVINO|sycl|SYCL).*

BuildSystem:	cmake
BuildOption:	-DCMAKE_C_COMPILER=clang
BuildOption:	-DCMAKE_CXX_COMPILER=clang++
BuildOption:	-DBUILD_SHARED_LIBS:BOOL=ON
BuildOption:	-DBARK_USE_SYSTEM_GGML:BOOL=ON
BuildOption:	-DENCODEC_USE_SYSTEM_GGML:BOOL=ON
BuildOption:	-DBARK_BUILD_EXAMPLES=%{build_examples}

%description
bark.cpp runs Suno Bark text-to-speech (and ships encodec.cpp for the
neural codec). Tensor kernels come from the system ggml package;
optional accelerators are separate:

* ggml-backend-blas — OpenBLAS
* ggml-backend-vulkan — Vulkan
* ggml-backend-opencl — OpenCL
* ggml-backend-hip — AMD ROCm/HIP

Convert official Bark checkpoints with convert.py (needs torch) to
a single ggml_weights.bin. CLI tools are in %{name}-examples, the
HTTP server in %{name}-server.

%package devel
Summary:	Development files for %{name}
Group:		Development/C++
Requires:	%{name}%{?_isa} = %{EVRD}
Requires:	cmake(ggml)

%description devel
Headers, pkg-config and CMake package config for bark.cpp
(libbark) and encodec.cpp (libencodec). Requires system ggml
(cmake(ggml)).

%if %{with examples}
%package server
Summary:	HTTP inference server for %{name}
Group:		Servers
Requires:	%{name}%{?_isa} = %{EVRD}
Recommends:	bark-cpp-model-bark-small-q4_0

%description server
HTTP server (bark-server) for bark.cpp. POST JSON {"text":"..."}
to /bark and get a WAV back.

Config: %{_sysconfdir}/sysconfig/bark-server
Unit:   bark.service  (default 127.0.0.1:8083, not upstream 1337,
        so it can run next to llama-server, whisper.service and
        stable-diffusion.service)

  systemctl enable --now bark.service

If MODEL is unset, the unit picks the first readable
ggml_weights.bin (or *.bin) under MODEL_DIRS
(/srv/ai then %{_datarootdir}/%{name}/models). Override with
MODEL=. There is no built-in API key. Do not bind HOST to a
public address without a reverse proxy.

  curl -s http://127.0.0.1:8083/bark \
    -H "Content-Type: application/json" \
    -d '{"text":"This is an audio generated by bark.cpp"}' \
    -o out.wav

%package examples
Summary:	CLI tools and helpers for %{name}
Group:		Sciences/Other
Requires:	%{name}%{?_isa} = %{EVRD}
Recommends:	bark-cpp-model-bark-small-q4_0

%description examples
CLI tools (bark-cli, bark-quantize, encodec-cli,
encodec-compress, encodec-decompress) and the convert.py /
download_weights.py helpers.

  bark-cli -m /srv/ai/bark-small/ggml_weights.bin \
    -p "This is an audio generated by bark.cpp" -o out.wav

  # CPU only (skip Vulkan / HIP plugins):
  bark-cli -ngl 0 -m /srv/ai/ggml_weights.bin -p "hello"

convert.py needs torch and huggingface-hub; it is not run at
build time. download_weights.py pulls the official Suno
checkpoints (network).
%endif

%prep
%setup -q -n bark.cpp-%{commit}
# Submodule checkout is empty in the GitHub archive.
rm -rf encodec.cpp
mkdir encodec.cpp
tar -xzf %{S:1} -C encodec.cpp --strip-components=1
rm -rf encodec.cpp/ggml encodec.cpp/.git encodec.cpp/.github
find . -name '.gitignore' -delete 2>/dev/null || true
%autopatch -p1

%if %{with examples}
%install -a
ln -sf bark-cli %{buildroot}%{_bindir}/bark

mkdir -p %{buildroot}%{_unitdir} %{buildroot}%{_sysconfdir}/sysconfig
mkdir -p %{buildroot}%{_libexecdir}/%{name}
cat >%{buildroot}%{_unitdir}/bark.service <<'UNIT'
[Unit]
Description=Text-to-speech HTTP server (bark.cpp)
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
# Wrapper sources the sysconfig (multiline MODEL_CANDIDATES).
# Do not use EnvironmentFile — systemd cannot parse a multiline list.
ExecStart=%{_libexecdir}/%{name}/bark-server-start
KillMode=process
Restart=on-failure
RestartSec=5s

# Unprivileged, no extra caps. Model file must be readable by this user
# (e.g. mode 0644 under /srv/ai — /home is hidden, see ProtectHome).
DynamicUser=yes
SupplementaryGroups=render video
UMask=0077
NoNewPrivileges=yes
CapabilityBoundingSet=
AmbientCapabilities=
RestrictSUIDSGID=yes
RestrictRealtime=yes
LockPersonality=yes
ProtectHostname=yes
ProtectClock=yes
ProtectKernelTunables=yes
ProtectKernelModules=yes
ProtectKernelLogs=yes
ProtectControlGroups=yes
ProtectProc=invisible
RestrictNamespaces=yes
RestrictAddressFamilies=AF_UNIX AF_INET AF_INET6 AF_NETLINK
SystemCallArchitectures=native
RemoveIPC=yes

# /usr /boot /etc read-only; /home /root invisible. Private /tmp.
# Models: /srv/ai is visible read-only ("-" = skip if missing).
# Extra trees: drop-in  ReadOnlyPaths=-/other/models
ProtectSystem=strict
ProtectHome=yes
PrivateTmp=yes
PrivateUsers=no
ReadOnlyPaths=-/srv/ai
# Mesa/Vulkan shader cache, HIP dumps. Server writes /tmp/bark_tmp.wav
# (PrivateTmp).
CacheDirectory=bark-server
Environment=XDG_CACHE_HOME=%{_localstatedir}/cache/bark-server

# GPU: DRM render node + KFD (ROCm). No MemoryDenyWriteExecute —
# Vulkan/HIP compile executable code.
DevicePolicy=closed
DeviceAllow=/dev/null rw
DeviceAllow=/dev/zero rw
DeviceAllow=/dev/urandom r
DeviceAllow=/dev/random r
DeviceAllow=char-drm rw
DeviceAllow=/dev/dri rw
DeviceAllow=/dev/kfd rw

[Install]
WantedBy=multi-user.target
UNIT

cat >%{buildroot}%{_libexecdir}/%{name}/bark-server-start <<'SH'
#!/bin/bash
# Resolve a Bark ggml_weights.bin and exec bark-server.
# Sourced config: /etc/sysconfig/bark-server
set -u

sysconfig=${BARK_SYSCONFIG:-/etc/sysconfig/bark-server}
# shellcheck source=/dev/null
[ -r "$sysconfig" ] && . "$sysconfig"

log() { printf 'bark-server: %s\n' "$*" >&2; }

readable() { [ -f "$1" ] && [ -r "$1" ]; }

# First readable path in a whitespace-separated list (globs allowed).
pick_from_candidates() {
	local cand f
	shopt -s nullglob
	for cand in $1; do
		case "$cand" in
		\#*) continue ;;
		esac
		for f in $cand; do
			if readable "$f"; then
				printf '%s\n' "$f"
				return 0
			fi
		done
	done
	return 1
}

# Prefer ggml_weights.bin, then any *.bin, in each MODEL_DIRS entry
# (including one-level subdirs such as bark-small/).
pick_from_dirs() {
	local dir f
	local -a dirs
	read -r -a dirs <<< "$1"
	shopt -s nullglob
	for dir in "${dirs[@]}"; do
		[ -z "$dir" ] && continue
		[ -d "$dir" ] || continue
		for f in \
			"${dir%/}/ggml-bark.bin" \
			"${dir%/}/ggml-bark-small.bin" \
			"${dir%/}/ggml-bark-small-q8_0.bin" \
			"${dir%/}/ggml-bark-small-q4_0.bin" \
			"${dir%/}/ggml_weights.bin" \
			"${dir%/}"/*/ggml_weights.bin \
			"${dir%/}"/*.bin
		do
			if readable "$f"; then
				printf '%s\n' "$f"
				return 0
			fi
		done
	done
	return 1
}

model=
if [ -n "${MODEL:-}" ]; then
	if readable "$MODEL"; then
		model=$MODEL
	else
		log "MODEL=$MODEL is not a readable file"
		exit 1
	fi
elif [ -n "${MODEL_CANDIDATES:-}" ]; then
	if ! model=$(pick_from_candidates "$MODEL_CANDIDATES"); then
		log "none of MODEL_CANDIDATES is readable"
		exit 1
	fi
else
	dirs=${MODEL_DIRS:-"/srv/ai /usr/share/bark-cpp/models"}
	if ! model=$(pick_from_dirs "$dirs"); then
		log "no ggml_weights.bin found under: $dirs"
		log "drop a converted Bark model in /srv/ai or set MODEL="
		exit 1
	fi
fi

log "using $model"
if [ -n "${BARK_PRINT_MODEL:-}" ]; then
	printf '%s\n' "$model"
	exit 0
fi

cmd=(/usr/bin/bark-server --model "$model")
[ -n "${HOST:-}" ] && cmd+=(--address "$HOST")
[ -n "${PORT:-}" ] && cmd+=(--port "$PORT")
if [ -n "${BARK_OPTIONS:-}" ]; then
	# Intentional split: BARK_OPTIONS is a flag string.
	# shellcheck disable=SC2206
	extra=( $BARK_OPTIONS )
	cmd+=("${extra[@]}")
fi

exec "${cmd[@]}"
SH
chmod 0755 %{buildroot}%{_libexecdir}/%{name}/bark-server-start

cat >%{buildroot}%{_sysconfdir}/sysconfig/bark-server <<'CFG'
# Bark ggml weights (convert.py → ggml_weights.bin).
# The unit hides /home and /root (ProtectHome=yes). Packaged files
# under /usr/share/bark-cpp/models are visible (ProtectSystem=strict
# leaves /usr readable). /srv/ai is extra (ReadOnlyPaths=-/srv/ai; "-"
# means the unit still starts if that directory is absent). Files must
# be readable by the service (e.g. chmod 0644). Extra trees: drop-in
#   ReadOnlyPaths=-/other/models
#
# Resolution, first match wins:
#   1. MODEL=            exact path (error if missing)
#   2. MODEL_CANDIDATES  whitespace-separated paths/globs
#   3. ggml_weights.bin, then */ggml_weights.bin, then *.bin
#      under each MODEL_DIRS entry
#MODEL=/srv/ai/bark-small/ggml_weights.bin
#MODEL_CANDIDATES="/srv/ai/bark-small/ggml_weights.bin /srv/ai/ggml_weights.bin"
MODEL_DIRS="/srv/ai /usr/share/bark-cpp/models"
# bark-server has no --api-key. Anyone who can reach HOST:PORT can
# synthesise speech. Keep HOST on loopback or put a reverse proxy
# with auth in front.
HOST=127.0.0.1
# 8083 so this can run next to llama-server (8080), whisper (8081)
# and stable-diffusion (8082). Upstream default is 1337.
PORT=8083
# GPU is on by default (ggml backends: Vulkan, HIP/ROCm, …).
# CPU only:             BARK_OPTIONS="-ngl 0"
# Thread count:         BARK_OPTIONS="-t 8"
BARK_OPTIONS=
CFG

mkdir -p %{buildroot}%{_datarootdir}/%{name}/models
install -m 0644 convert.py download_weights.py requirements.txt \
	%{buildroot}%{_datarootdir}/%{name}/
install -m 0644 encodec.cpp/convert.py \
	%{buildroot}%{_datarootdir}/%{name}/encodec-convert.py
install -m 0644 README.md \
	%{buildroot}%{_datarootdir}/%{name}/
%endif

%files
%license LICENSE
%{_libdir}/libbark.so.*
%{_libdir}/libencodec.so.*

%files devel
%doc README.md
%{_includedir}/bark.h
%{_includedir}/encodec.h
%{_libdir}/libbark.so
%{_libdir}/libencodec.so
%{_libdir}/cmake/bark/
%{_libdir}/cmake/encodec/
%{_libdir}/pkgconfig/bark.pc
%{_libdir}/pkgconfig/encodec.pc

%if %{with examples}
%files server
%{_bindir}/bark-server
%{_libexecdir}/%{name}/bark-server-start
%{_unitdir}/bark.service
%config(noreplace) %{_sysconfdir}/sysconfig/bark-server

%files examples
%{_bindir}/bark
%{_bindir}/bark-cli
%{_bindir}/bark-quantize
%{_bindir}/encodec-cli
%{_bindir}/encodec-compress
%{_bindir}/encodec-decompress
%exclude %{_bindir}/bark-server
%{_datarootdir}/%{name}/
%endif