ArchGPU

Multi-vendor GPU, bootloader, Wayland, and gaming-stack configurator for Arch Linux. One-click hardware scan; auto-applies only the tweaks your host actually needs; verifies every change is live in the running kernel before marking it done.

Rust + Slint. GUI + CLI. pacman-packaged. Idempotent. State-aware.

Formerly arch-nvidia-tweaker. Renamed to reflect reality: the tool now drives NVIDIA, AMD (amdgpu), and Intel (i915/xe) equally — plus bootloader cmdline injection across GRUB, systemd-boot, Limine, and UKI, plus AUR helper bootstrap, plus legacy-config sanitation.

---

Install

One command (recommended)

git clone https://github.com/archledger/archgpu.git
cd archgpu
./install.sh

install.sh does everything:

1. Installs base-devel if missing (fakeroot + make + gcc — makepkg prerequisites)
2. Installs rust if missing (compiler toolchain)
3. Builds and installs the pacman package via makepkg -si

Safe to re-run: each step checks before acting.

If you'd rather do it manually

sudo pacman -S --needed base-devel rust
git clone https://github.com/archledger/archgpu.git
cd archgpu/packaging
makepkg -si

If you've already built and just want to install the .pkg.tar.zst

cd packaging
sudo pacman -U archgpu-*-x86_64.pkg.tar.zst

---

What it does

Action	Scope
GPU detection	Enumerates every display controller via lspci, classifies vendor (NVIDIA / AMD / Intel) and kernel driver (xe vs i915, amdgpu vs radeon). Recommends the correct driver package per GPU architecture (NVIDIA: nvidia-open-dkms for Turing+, AUR nvidia-580xx-dkms for Maxwell/Volta/Pascal, AUR nvidia-470xx-dkms for Kepler, AUR nvidia-390xx-dkms for Fermi).
Kernel cmdline	GPU-aware, multi-bootloader. Writes nvidia-drm.modeset=1 + nvidia-drm.fbdev=1 (NVIDIA), amdgpu.ppfeaturemask=0xffffffff (amdgpu), i915.enable_guc=3 (i915). xe gets nothing — it handles GuC/HuC natively. Supports GRUB (grub-mkconfig), systemd-boot (bootctl update), Limine (no regen needed), UKI (mkinitcpio -P).
Live verification	After writing, every cmdline param is verified against /sys/module/<m>/parameters/<p>. UI shows ✓ Active (green) only when the running kernel confirms; ⟳ Reboot pending (yellow) when config is written but kernel hasn't adopted it yet.
Wayland	Minimal /etc/profile.d/99-nvidia-wayland.sh (comment-only — legacy GBM_BACKEND=nvidia-drm global export is deliberately NOT set; it breaks hybrid PRIME); initramfs modules via MODULES+=() drop-in (never touches HOOKS — removing the kms hook breaks simpledrm console handoff); PRIME OutputClass Xorg drop-in for hybrid hosts.
Power	nvidia-suspend / nvidia-hibernate / nvidia-resume systemd units, NVreg_UseKernelSuspendNotifiers (all NVIDIA hosts), NVreg_DynamicPowerManagement=0x02 (laptops only), nouveau blacklist drop-in.
Essentials (Phase 26)	Vendor-agnostic userspace baseline: vulkan-icd-loader + vulkan-tools + clinfo + libva-utils + vdpauinfo on every host; mesa + lib32-mesa + vendor ICDs (vulkan-radeon / vulkan-intel) on AMD / Intel; split firmware (linux-firmware-amdgpu / linux-firmware-intel — carved out of linux-firmware in 2025); VA-API driver generation-gated on Intel (intel-media-driver for Gen 8+, libva-intel-driver for Gen 6/7). Non-gaming; run via --apply-essentials.
Gaming	[multilib] uncomment; per-vendor Vulkan (RADV for AMD, vulkan-intel for Intel, nvidia-utils + libva-nvidia-driver for NVIDIA); gamemode + mangohud + lib32 variants; vm.max_map_count=1048576 sysctl; legacy NVIDIA drivers bootstrapped from AUR via yay (which is itself auto-installed if missing).
Legacy sanitation	Warns the user — in the GUI and in --diagnose — about configs that follow outdated tutorials: AMDVLK alongside RADV, deprecated xf86-video-intel DDX, mesa-vdpau / libva-mesa-driver (both defunct — Mesa 26 bundles VA-API in-tree), manually-authored /etc/X11/xorg.conf, WLR_NO_HARDWARE_CURSORS=1 or GBM_BACKEND=nvidia-drm in other /etc/profile.d/ scripts.
SUDO_ASKPASS	yay's internal sudo is routed through a DE-appropriate graphical askpass (ksshaskpass, seahorse-ssh-askpass, lxqt-openssh-askpass) so AUR installs don't hang the GUI when no TTY is available.

---

Usage

GUI

Launch ArchGPU from your application menu (pkexec prompts for auth; polkit preserves your Wayland/X11 env so the window renders). Workflow:

1. Review the Detected hardware card — chassis, GPU inventory, bootloader type.
2. Click ✨ Auto-Optimize — the hero button flips on exactly the switches that would improve this host AND aren't already active/pending.
3. Leave Dry run on for Preview; turn it off to actually Apply.
4. Watch the live pacman / mkinitcpio / yay / grub-mkconfig / bootctl stream in the dark terminal-style Console card.

CLI

# Read-only
archgpu --detect                  # hardware + bootloader + per-tweak recommendation
archgpu --diagnose                # 14-point issue scan with remediation hints

# Preview (no writes, no subprocess side-effects)
archgpu --dry-run --apply-all

# Apply individual areas
sudo archgpu --apply-groups            # add invoking user to video + render (re-login to activate)
sudo archgpu --apply-essentials --yes  # Mesa + Vulkan loader + split firmware + VA-API + diag userspace
sudo archgpu --apply-wayland           # env drop-in + initramfs modules + PRIME (hybrid)
sudo archgpu --apply-bootloader        # GPU-aware cmdline + per-bootloader regeneration
sudo archgpu --apply-power             # suspend services + modprobe + nouveau blacklist
sudo archgpu --apply-gaming --yes      # multilib + Vulkan + gamemode + mangohud (+ AUR if needed)

# Reverse cleanup (Phase 28) — destructive, opt-in only, NOT in --apply-all
sudo archgpu --dry-run --apply-cleanup    # PREVIEW: list what would be removed and why
sudo archgpu --apply-cleanup --yes        # actually remove (writes pre-cleanup snapshot first)

# Smart troubleshoot (Phase 29) — detect → fix → verify per recipe
sudo archgpu --dry-run --apply-troubleshoot   # detect-only; explain causes, no writes
sudo archgpu --apply-troubleshoot             # apply auto-fixable recipes + verify
# Recipes (initial set): nomodeset_stuck, nouveau_active_with_nvidia,
#                        dangling_vulkan_icd, software_rendering (diagnostic-only)

# All at once (does NOT include --apply-cleanup or --apply-troubleshoot;
# pass them explicitly to combine)
sudo archgpu --apply-all --yes

--yes passes --noconfirm to pacman (required for GUI installs with no TTY).

---

UI state badges

Every per-tweak row shows exactly one of:

Badge	Meaning
(Switch)	Unapplied — click Apply to make the change
✓ Active (green)	Config written AND the running kernel confirms it
⟳ Reboot (yellow)	Config written; kernel hasn't adopted it yet
Unsupported (orange)	Doesn't apply to this host (e.g. NVIDIA tweak on Intel-only box)

The live-kernel probe reads /sys/module/nvidia_drm/parameters/{modeset,fbdev}, /sys/module/amdgpu/parameters/ppfeaturemask, and /sys/module/i915/parameters/enable_guc for the bootloader tweak — so green means "running right now", not "we edited a text file and hoped for the best".

---

Architecture

Module	Responsibility
src/core/bootloader.rs	BootloaderType + detect_active_bootloader + per-type apply_{uki,grub,sdb,limine}, pure param-parsers (grub_add_param / sdb_add_param / limine_add_param), live-kernel probes (/sys/module/<m>/parameters/<p>), 2-stage check_state → Active / PendingReboot / Unapplied / Incompatible
src/core/gpu.rs	lspci -mm -nn -D parser, GpuVendor / NvidiaGeneration classification, driver-family helpers (uses_xe_driver, uses_i915_driver, uses_amdgpu_driver, uses_radeon_legacy_driver), NVIDIA driver package recommendation per arch
src/core/hardware.rs	SMBIOS chassis_type → FormFactor::{Laptop, Desktop, Unknown}
src/core/wayland.rs	Modern comment-only profile.d drop-in, MODULES+=() mkinitcpio drop-in (HOOKS untouched), hybrid PRIME Xorg config, sanitation scanner for /etc/X11/xorg.conf + WLR_NO_HARDWARE_CURSORS + global GBM_BACKEND=nvidia-drm
src/core/power.rs	modprobe drop-in (universal + laptop-specific options), nouveau blacklist, systemctl enable of nvidia-suspend/hibernate/resume
src/core/essentials.rs (Phase 26)	Vendor-agnostic userspace baseline: ALWAYS_ON_PACKAGES (Vulkan loader + diag tools) + vendor-conditional Mesa / RADV / ANV / split firmware / VA-API driver. Intel VA-API routing is device-ID-gated (< 0x1600 → libva-intel-driver, else intel-media-driver).
src/core/gaming.rs	[multilib] state-machine uncommenter, GPU-aware package resolver (resolve_gaming_packages + resolve_aur_packages), sanitation scanner for AMDVLK / xf86-video-intel / mesa-vdpau / libva-mesa-driver, vm.max_map_count sysctl
src/core/groups.rs (Phase 27)	Provision invoking user's membership in video + render via usermod -aG. Re-login (NOT reboot) required for the change to reach the current session — surfaced in the Apply-time detail message. Pure parser for /etc/group + getent-style runtime probe.
src/core/aur.rs	Helper detection (yay / paru), invoking_user via SUDO_USER / PKEXEC_UID + allowlist, manual yay-bin bootstrap (git clone + makepkg as user → pacman -U as root), SUDO_ASKPASS routing per DE
src/core/prime.rs	Xorg OutputClass drop-in for hybrid GPUs (skipped when nvidia-utils ships its own)
src/core/diagnostics.rs	14-point read-only scanner, Finding{severity,title,detail,fix_hint}, surfaces gaming + wayland sanitation warnings
src/core/cleanup.rs (Phase 28)	Reverse-cleanup engine: compute_removal_plan(gpus, installed) classifies each candidate as HardwareAbsent / Defunct / LegacyDeprecated / ConflictingChoice. apply() writes a pre-cleanup pacman -Qq snapshot to /var/backups/archgpu/pre-cleanup-<ts>.txt before any pacman -Rns. NEVER_REMOVE allowlist as defense-in-depth. Opt-in only — never in Actions::all() or auto::recommend.
src/core/troubleshoot.rs (Phase 29)	Recipe trait + initial recipes (nomodeset_stuck, nouveau_active_with_nvidia, dangling_vulkan_icd, software_rendering). Each runs detect → cause → fix → verify. Verification::{LiveVerified, PendingReboot, Failed, NotApplicable} distinguishes "fix took effect now" from "config written, reboot to confirm" from "fix attempted but didn't help." Opt-in only — never in Actions::all() or auto::recommend.
src/core/auto.rs	recommend(ctx, form, gpus) -> Actions — cross-references hardware applicability AND check_state().is_unapplied(); used both at GUI startup and by the Auto-Optimize hero
src/core/state.rs	TweakState::{Active, PendingReboot, Unapplied, Incompatible} + helpers
src/utils/process.rs	run_streaming(cmd, on_line) — pipes stdout+stderr through mpsc into an FnMut(&str)
src/utils/fs_helper.rs	atomic_write (temp + fsync + rename), backup_to_dir (timestamped /var/backups/archgpu/), write_dropin, ChangeReport::{AlreadyApplied, Applied, Planned}
src/cli.rs	clap CLI; dispatches into core::run_actions with a println! progress sink
src/gui.rs	Slint wiring; populate_detection polls every check_state; append_log updates log-text + triggers reactive auto-scroll in the Console card
ui/main_window.slint	Dark Libadwaita-style UI: Card / HeroButton / SecondaryButton / StateBadge / TweakRow components
packaging/PKGBUILD	Builds directly from project root ($startdir/..) — never touches makepkg's $srcdir. options=('!debug') so debugedit isn't required.

---

Tests

cargo test --release --locked

110+ passing tests covering GPU parsing across real device IDs, NVIDIA architecture classification, per-vendor package resolution, multilib state-machine editing, bootloader detection for all four types with UKI+SDB coexistence, per-bootloader param-add helpers (quoted / unquoted / empty / idempotent / commented-line edge cases), check_state transitions (Active / PendingReboot / Unapplied / Incompatible) including live-sysfs seeding, Auto-Optimize recommendation rules across five hardware scenarios, atomic_write / backup_to_dir / write_dropin behavior, subprocess streaming (interleaved stdout/stderr, non-zero exit propagation), sanitation matrix (AMDVLK / xf86-video-intel / mesa-vdpau / legacy X11 configs / deprecated env vars).

cargo clippy --all-targets --all-features -- -D warnings
# 0 warnings, fails on any

---

Runtime dependencies

• pciutils (lspci)
• polkit (pkexec launcher)
• Slint Linux graphics stack: fontconfig, freetype2, libxkbcommon, wayland, libx11, libxcb, libxcursor, libxi, libxrandr, mesa

optdepends

• nvidia-utils, nvidia-prime — runtime userspace + prime-run
• gamemode, mangohud — auto-installed by --apply-gaming
• yay — auto-bootstrapped if missing and a legacy AUR driver is needed
• sbctl — Secure Boot signing (via its own mkinitcpio post-hook; this tool doesn't sign)
• ksshaskpass (KDE/Wayland) or seahorse (GNOME) — graphical askpass for yay's inner sudo

---

Safety

• Every mutating write is preceded by a timestamped backup in /var/backups/archgpu/.
• Every write is atomic (temp file + fsync + rename). Crash-safe.
• --dry-run runs the full detection + planning pipeline with zero side effects.
• NVIDIA-specific actions are hard-gated on NVIDIA GPU presence. On AMD/Intel hosts they cleanly report "skipped" instead of doing damage.
• yay is never run as root. The bootstrap install(1)s a user-owned /tmp dir, then sudo -u <user> for git clone + makepkg, finishing with pacman -U as root — no user sudo prompt is required for the bootstrap itself.
• $SUDO_USER / $PKEXEC_UID→getent are validated against a POSIX allowlist ([A-Za-z0-9_-]{1,32}) before interpolation into any path or subprocess.
• Master configs (/etc/default/grub, /etc/pacman.conf, /etc/kernel/cmdline, /boot/loader/entries/*.conf, /boot/limine.conf) are only touched when the config system provides no drop-in alternative — and those edits are surgical (preserve surrounding content) and backed up first. Everything else (profile.d, modprobe.d, mkinitcpio.conf.d, sysctl.d, X11/xorg.conf.d) uses drop-ins.
• /etc/mkinitcpio.conf is never touched directly. We write only to /etc/mkinitcpio.conf.d/nvidia-modules.conf with MODULES+=(…). The HOOKS=() array is never modified — removing kms would break simpledrm console handoff.

---

License

Dual-licensed: MIT OR Apache-2.0 (same as most Rust crates).