OPINIONATED OMARCHY
TROUBLESHOOTING CORPUS
790
RECORDS
790
AUDITED
0
UNCHECKED
2388
SOURCES
LIVE
gpu-drivers high common DISPUTE

Fix corrupted browser video and a dGPU that never sleeps on a hybrid laptop whose panel runs on the iGPU

omarchyhyprlandnvidiaintelamdlaptopchromiumchromeedgeelectronteams
CORRECTED An auditor found the fix wrong or incomplete and rewrote it.
AUDIT NOTE

Confirmed on this machine, omarchy 4.0.2-1: /usr/share/omarchy/default/hypr/nvidia.lua sets all three variables behind omarchy-hw-nvidia plus omarchy-hw-nvidia-gsp with no display check, /usr/share/omarchy/default/hypr/envs.lua requires default.hypr.nvidia, /usr/share/omarchy/default/hypr/omarchy.lua:16 requires default.hypr.envs, and /usr/share/omarchy/default/hypr/autostart.lua:3-4 runs systemctl --user import-environment and dbus-update-activation-environment --systemd --all. The file is byte-identical to the v4.0.3 tag fetched with gh api, so claim 1 holds, and I also ran the three detectors here: nvidia and nvidia-gsp both succeed and the session really carries LIBVA_DRIVER_NAME=nvidia, NVD_BACKEND=direct and __GLX_VENDOR_LIBRARY_NAME=nvidia. Claim 2 holds: nothing under /usr/share/omarchy requires hypr.envs, the only match in the tree is default.hypr.envs, and the shipped config/hypr/hyprland.lua requires hypr.monitors, hypr.input, hypr.bindings, hypr.looknfeel and hypr.autostart and no envs file, so ~/.config/hypr/envs.lua is dead, which is upstream issue 9902. Claim 3 holds by load order in that same file, where require("default.hypr.omarchy") runs above the personal requires, so an hl.env placed before it loses. Claim 4, Q7, is honest: issue 8989 carries the Quadro P520 report with nvidia_uvm used_by=0 and no LIBVA_DRIVER_NAME on that branch, no confirmed fix, and only an untested __EGL_VENDOR_LIBRARY_FILENAMES hypothesis, but the record says "Pascal and older" where omarchy-hw-nvidia-without-gsp is bounded 0x1340 to 0x1dff, so Maxwell, Pascal and Volta take that branch and anything older than Maxwell matches neither branch and gets no NVIDIA environment at all. Two defects, both in the gate rather than the mechanism. First, the fix tells the reader to apply the override only when every NVIDIA connector is disconnected and disabled and the danger says the shipped setting is correct otherwise, and I read four machines in issue 8215 that falsify this: Hakira-Shymuy, RTX 4060 driving HDMI-A-1 with a monitor attached while Brave's GPU process stays on the AMD node, marioxabel, RTX 2060 driving DP-3 while Chromium stays on the Intel node, jhostileo, fixed with an hl.env override on an AMD panel plus NVIDIA HDMI plus dock, and a Lenovo i7-13800H with an RTX 4060 on DP-2 whose single-variable A/B gives a blank YouTube window under nvidia and clean playback under iHD. Scanout ownership is not render-device ownership, triage on issue 10410 settled the same point, so I rewrote the gate around which render node the compositor holds and rewrote the danger to match. Second, the fix says the __GLX_VENDOR_LIBRARY_NAME line was confirmed by an Arrow Lake plus RTX 5070 reporter, and that confirmation does not exist: that reporter set LIBVA_DRIVER_NAME only and pointed out that __GLX survives it, the issue 9890 reporter wrote that he had not tested __GLX, and the Arrow Lake plus RTX 5070 Ti reporter who did set both saw the eglCreateImage storm stop while the corruption persisted without a full logout. Everything else held from source: the CHECK_EQ modifier isolation in issue 9890, the decode-only entrypoint tables, the 99.9 Wh one-third idle draw measurement in issue 8989's body, the 4.0.1 attribution to 33d7363c which gh api shows inside the v4.0.0..v4.0.1 range, and pull requests 7851, 9483 and 10588 all open on quattro as of 2026-09-11. Not exercised: this is a desktop with one NVIDIA card driving card1-HDMI-A-1 and renderD128 on vendor 0x10de, so the hybrid path, the corruption, the crash and the power behaviour could not be reproduced here, and all of those rest on the reporters.

The cause above was rewritten on 2026-09-11 to match this note; the fix was corrected by the audit itself.
RISK

On a machine where the NVIDIA card is the rendering device, a desktop or a muxed laptop in dGPU mode, LIBVA_DRIVER_NAME=nvidia matches the render device and this override will break hardware decode instead of fixing it. Check which render node the compositor and the browser hold first, as the fix shows. A connected NVIDIA output is not that test: on a hybrid laptop the dGPU can scan out an external monitor while the compositor and the browser still render on the iGPU, and the override applies there too.

Symptom

Omarchy 4.0.1 or later on a hybrid laptop whose Intel or AMD iGPU renders the session. Usually the NVIDIA card drives nothing, but the same failure happens when it scans out an external monitor while the compositor and the browser still render on the iGPU.

Chromium, Chrome, Edge, Discord, Teams and other Electron apps show corrupted or black video, ghosting while scrolling, elements that paint only on mouse hover, hard scroll stalls on video-heavy pages, or a WebRTC stream that never paints. Live streams with a chat panel open are the most reliable trigger, while static pages stay smooth and ordinary playback can look fine. Rarely the browser's GPU process aborts with SIGTRAP and leaves a chromium core with --type=gpu-process in coredumpctl. Launched from a terminal, Chromium prints hundreds of:


eglCreateImage failed with 0x00003009
Unable to initialize binding from pixmap
OzoneImageBacking::ProduceSkiaGanesh failed to create GL representation

The session carries the NVIDIA values even though the NVIDIA card is not the rendering device:


$ systemctl --user show-environment | grep -E 'LIBVA|NVD_BACKEND|__GLX'
LIBVA_DRIVER_NAME=nvidia
NVD_BACKEND=direct
__GLX_VENDOR_LIBRARY_NAME=nvidia

vainfo reports the NVDEC driver rather than the Intel or AMD one, and installing intel-media-driver alone changes nothing. On most of the reported machines every NVIDIA connector under /sys/class/drm/ reads disconnected and disabled while the iGPU's eDP-1 is connected, but on several the NVIDIA card has a live external output and the browser's GPU process is still pinned to the iGPU's render node.

On battery the dGPU never runtime-suspends while a browser is open. /sys/bus/pci/devices/<dgpu>/power/runtime_status stays active and runtime_suspended_time stays 0. One reporter measured the dGPU at about a third of idle draw on a 99.9 Wh laptop.

Cause

Established in the threads by the repository's triage collaborator, from source, and confirmed on an omarchy 4.0.2-1 workstation where nvidia.lua and all three detectors read exactly as described. The file is unchanged at the v4.0.3 tag. The cluster's canonical thread is issue 8215.

/usr/share/omarchy/default/hypr/nvidia.lua, loaded from default/hypr/envs.lua, which default/hypr/omarchy.lua requires, runs hl.env("NVD_BACKEND", "direct"), hl.env("LIBVA_DRIVER_NAME", "nvidia") and hl.env("__GLX_VENDOR_LIBRARY_NAME", "nvidia") whenever bin/omarchy-hw-nvidia and bin/omarchy-hw-nvidia-gsp both succeed. Those detectors only check that a PCI device with vendor 0x10de at class 0x03* exists with a device id at or above 0x1e00, which is Turing and newer. Neither asks whether the NVIDIA GPU renders or scans out anything, and class 0x0302, a 3D controller, drives no display by definition. default/hypr/autostart.lua then exports the whole environment session-wide on hyprland.start through systemctl --user import-environment and dbus-update-activation-environment --systemd --all, so every app inherits it.

LIBVA_DRIVER_NAME overrides libva's DRM-based autodetection, so every VA-API client loads nvidia_drv_video.so against the iGPU's render node. Compositing then happens on the iGPU while NVDEC decode runs on the NVIDIA card and the exported frames cross GPUs, which fails on import with EGL_BAD_MATCH. A contributor isolated the rare abort to Chromium's CHECK_EQ(handle.modifier, object.drm_format_modifier) in VaapiWrapper::ExportVASurfaceAsNativePixmapDmaBufUnwrapped, because libva-nvidia-driver returns a different NVIDIA block-linear modifier for the half-height UV plane than for the Y plane. The NVIDIA VA driver is decode only and exposes no VAEntrypointEnc* entrypoints, so the pin removes the iGPU's hardware encode as well. Any app touching video decode also opens /dev/nvidia-uvm, which is what blocks runtime suspend. Which mechanism breaks Teams was not isolated.

Scanout is not the same thing as rendering, which is why "does an NVIDIA connector report connected" is the wrong question to ask about a machine. Four machines in the tracker have a live NVIDIA output and still show the bug, because the compositor keeps rendering on the iGPU while the dGPU only scans out the ports wired to it, and the browser follows the compositor's render node. What decides it is the render node the compositor and the browser actually open.

The branch went live in 4.0.1 through commit 33d7363c, which fixed o.shell_succeeds() always returning false inside Hyprland and so un-masked code that had been dead in 4.0.0. These machines were accidentally correct before that.

Two limits established in the threads. An hl.env placed before require("default.hypr.omarchy") loses, because the default loads after it, and the shipped config/hypr/hyprland.lua says plainly that the personal files load after the defaults. And ~/.config/hypr/envs.lua is never loaded by anything, confirmed here: nothing in /usr/share/omarchy requires hypr.envs, the only match in the tree is default.hypr.envs at default/hypr/omarchy.lua:16, so an override written there does nothing. That is issue 9902.

Maxwell, Pascal and Volta take the other branch, device ids from 0x1340 to 0x1dff, which sets NVD_BACKEND=egl and __GLX_VENDOR_LIBRARY_NAME=nvidia with no LIBVA_DRIVER_NAME. Anything older than Maxwell matches neither detector and gets no NVIDIA environment at all. One Quadro P520 reporter found the dGPU pinned awake with nvidia_uvm used_by=0 and no process holding a CUDA context, so the mechanism above is not what holds it there, and that machine is not fixed by the override below. None of the open pull requests fixes it either, because all of them work on the GSP branch. The thread had no confirmed fix for it, only an untested suggestion of __EGL_VENDOR_LIBRARY_FILENAMES=/usr/share/glvnd/egl_vendor.d/50_mesa.json.

Three upstream pull requests propose different gates, 7851, 9483 and 10588, and none has merged as of 2026-09-11. Triage's recommendation on issue 10410 is 7851, which reads PCI vendors rather than connectors and so cannot flip when a monitor is docked, with one condition: a fix has to clear LIBVA_DRIVER_NAME rather than merely decline to set it, because systemctl --user import-environment only adds or overwrites names and anyone who has run 4.0.1 or 4.0.2 already has the stale value in their systemd user environment.

Fix

Omarchy 4 only, and only on a machine where the iGPU is the rendering device. Check that first, because where the NVIDIA card renders, a desktop or a muxed laptop in dGPU mode, the shipped setting matches the render device and this override is wrong.

Find which render node the compositor holds, and which PCI vendor owns it:


ls -l /proc/$(pgrep -x Hyprland | head -1)/fd | grep -oE 'renderD[0-9]+' | sort | uniq -c
for n in /sys/class/drm/renderD*; do
  printf '%s %s vendor=%s\n' "$(basename "$n")" "$(basename "$(readlink -f "$n/device")")" "$(cat "$n/device/vendor")"
done

Vendor 0x8086 is Intel, 0x1002 is AMD, 0x10de is NVIDIA. If the node the compositor holds belongs to the Intel or AMD iGPU, the override applies. Check the browser the same way once it is running, because it picks its own node:


ls -l /proc/$(pgrep -f 'type=gpu-process' | head -n1)/fd | grep -oE 'renderD[0-9]+' | sort | uniq -c

Connector state is a useful cross-check and nothing more. A dGPU with a monitor plugged into one of its own ports can still be a machine whose compositor and browser render on the iGPU, and four machines in the upstream cluster are exactly that, so do not read a connected NVIDIA output as "the shipped setting is right here":


for c in /sys/class/drm/card*-*; do printf '%s status=%s enabled=%s\n' "$(basename "$c")" "$(cat "$c/status")" "$(cat "$c/enabled")"; done

Intel iGPU, install the media driver first:


omarchy update                      # fetch the sync databases first on a fresh install
sudo pacman -S intel-media-driver

Then override the variable from your own config, which loads after Omarchy's defaults, so the last hl.env assignment wins. Put it in ~/.config/hypr/hyprland.lua **below** the require("default.hypr.omarchy") line:


-- ~/.config/hypr/hyprland.lua, below require("default.hypr.omarchy")
hl.env("LIBVA_DRIVER_NAME", "iHD")              -- Intel iGPU
hl.env("__GLX_VENDOR_LIBRARY_NAME", "mesa")

On an AMD iGPU use radeonsi in place of iHD. Mesa is already installed there. ~/.config/hypr/autostart.lua works as well, because hyprland.lua requires it after the defaults, and one reporter fixed an AMD machine that way. Do not use ~/.config/hypr/envs.lua, which nothing loads.

Both lines together are the workaround the issue 8989 reporter validated, and after a relogin he reported no corruption, the dGPU in D3cold even during playback, and about 5 W less idle draw. The LIBVA_DRIVER_NAME line is the one that carries the fix: two other reporters cleared their machines with it alone, one on an Arrow Lake plus RTX 5070 Max-Q laptop who then removed --disable-gpu-compositing and kept full GPU acceleration, and one who restored incoming Discord streams with iHD. A third set both lines on an Arrow Lake plus RTX 5070 Ti, saw the eglCreateImage errors stop, and still had corruption in the live session because he did not log out, so do not skip the relogin below.

Then log out and back in. hyprctl reload re-parses the config and processes Hyprland spawns afterwards do see the new value, but Omarchy's autostart runs systemctl --user import-environment at session start, so apps launched through systemd user scopes keep the old value until the next login. Fully quit any browser or Electron app that was running before the change, because each keeps its old environment until restarted.

Leave NVD_BACKEND=direct alone. The same branch still sets it and it is inert once the NVIDIA VA driver is no longer loaded.

Remove any --disable-gpu-compositing line from ~/.config/chromium-flags.conf or chrome-flags.conf afterwards, because reporters found full GPU compositing works once the driver matches. --disable-features=VaapiVideoDecoder does not fix this, because the NVIDIA VA driver still loads.

Apps that should run on the dGPU keep working with PRIME offload scoped to their own launch:


__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia <game>

Verify

After a fresh login:


systemctl --user show-environment | grep -E 'LIBVA|__GLX'   # iHD or radeonsi, and mesa
vainfo 2>/dev/null | head -3                                # the iGPU driver, not NVDEC
dgpu=$(grep -l 0x10de /sys/bus/pci/devices/*/vendor | head -1 | xargs dirname)
cat "$dgpu/power/runtime_status"                            # suspended, even during playback
cat "$dgpu/power/runtime_suspended_time"                    # grows over time

Then restart the browser and check what its GPU process actually maps:


GP=$(pgrep -f 'type=gpu-process' | head -n1)
grep -oE '/usr/lib/dri/[a-z0-9_]+_drv_video\.so' /proc/$GP/maps | sort -u
grep -oE 'lib(cuda|nvcuvid)\.so[.0-9]*' /proc/$GP/maps | sort -u
ls -l /proc/$GP/fd | grep -oE 'renderD[0-9]+|nvidia[a-z0-9-]*' | sort | uniq -c

On Intel, iHD_drv_video.so is mapped, the libcuda and libnvcuvid grep is empty, and the /dev/nvidia0 and /dev/nvidia-uvm fds are gone. On AMD, nvidia_drv_video.so is absent, since the radeonsi driver resolves to libgallium-*.so, and the count of /dev/nvidia0 fds is zero. Do not read /proc/<pid>/environ of the browser process, because Chromium rewrites that region for process titles and it reads back as garbage. The chrome_crashpad_handler child inherits the same environment and reads correctly. Chromium should also play video with no eglCreateImage failed with 0x00003009 lines.

Sources

nvidia-env-forced-on-igpu-primary-hybrid-laptop