Frank
03d44d3308
ComfyUI had failed 9,146 times and had been down ~7 hours when this was found:
its container was created 2026-09-15 14:12, BEFORE the GPU declarations were
added to the compose service, and `restart: unless-stopped` kept restarting
that stale container — no /dev/kfd, no /dev/dri, no keep-groups — so it died
every time on `RuntimeError: No HIP GPUs are available`. Its "unit" was a
hand-written ~/.config/systemd/user/comfyui.service that shelled out to
`podman compose up -d`, which is not repo-tracked state and collides by name
with the Quadlet the generator produces.
- dankclaw/comfyui/quadlet/comfyui.container: image comfyui-rocm:gfx1151
(built from Containerfile.comfyui), /dev/kfd + /dev/dri, keep-groups,
the four comfyui-data volumes, loopback-only publish (an unguarded service
is internal-only here; the homepage card probes 127.0.0.1:8188), Restart=
always, WantedBy=default.target.
- the compose service is deleted — one owner per container name.
- deploy_comfyui builds the image, installs the unit, retires the compose-era
unit into ~/.AGENTTRASH, removes the stale container so the unit's own can
take the name, and waits on :8188.
- the healthcheck is a file, because the podman 5.x v1 generator splits
HealthCmd on spaces and an inline `python3 -c "..."` becomes a quoting
puzzle that fails silently.
The comfyui seam now reads the unit instead of compose for devices, command and
groups (and resolves the Containerfile from the deployer's declaration, not
from a directory listing).
97 lines
5.1 KiB
Bash
97 lines
5.1 KiB
Bash
#!/usr/bin/env bats
|
|
# Seam: ComfyUI ROCm GPU configuration on Dankclaw.
|
|
# The gfx1151-enabled PyTorch build is the only verified GPU path on Strix
|
|
# Halo. The compose must:
|
|
# - build from a Containerfile based on the gfx1151 ROCm torch image
|
|
# - pass /dev/kfd and /dev/dri to the container
|
|
# - NOT pass --cpu
|
|
# - grant GPU access via world-writable compute nodes (udev MODE 0666), not
|
|
# named-group membership, because systemd user-unit spawns carry NO
|
|
# supplementary groups at spawn time
|
|
#
|
|
# STATUS (2026-09-16): ComfyUI is a Quadlet unit, not a compose service. The
|
|
# build still comes from Containerfile.comfyui (AMD's gfx1151 torch base) via
|
|
# the deployer's build_compose_image, and the wiring now lives in
|
|
# dankclaw/comfyui/quadlet/comfyui.container: /dev/kfd + /dev/dri, keep-groups
|
|
# only, and the udev rule at dankclaw/udev/99-comfyui-gpu-access.rules (MODE
|
|
# 0666 on kfd + renderD*) for compute-node access, because systemd user-unit
|
|
# spawns carry no supplementary groups. The move happened after the compose
|
|
# container went stale — created before the GPU declarations existed, then
|
|
# restarted 9,146 times without them. See
|
|
# serverAppConfigs-docs/Research/ComfyUI ROCm and Vulkan.md.
|
|
|
|
load helpers
|
|
|
|
@test "comfyui builds from the gfx1151 ROCm Containerfile" {
|
|
# The image is built by the deployer (build_compose_image comfyui, which
|
|
# resolves dankclaw/Containerfile.comfyui) and referenced by the unit.
|
|
grep -q "comfyui)" "$REPO_ROOT/scripts/deploy-remote.sh"
|
|
grep -q 'image=comfyui-rocm:gfx1151' "$REPO_ROOT/scripts/deploy-remote.sh"
|
|
grep -q 'containerfile=dankclaw/Containerfile.comfyui' "$REPO_ROOT/scripts/deploy-remote.sh"
|
|
local unit="$REPO_ROOT/dankclaw/comfyui/quadlet/comfyui.container"
|
|
grep -q '^Image=localhost/comfyui-rocm:gfx1151$' "$unit" \
|
|
|| { echo "unit must run the locally built image"; return 1; }
|
|
# One owner per container name: a surviving compose service would fight the
|
|
# unit for the container, which is how the stale one stayed alive.
|
|
! grep -qE '^\s+comfyui:' "$REPO_ROOT/dankclaw/compose.yaml"
|
|
}
|
|
|
|
@test "comfyui Containerfile is based on the gfx1151 torch image" {
|
|
# Resolve the Containerfile from the deployer's declaration — the file the
|
|
# build actually uses — not from the filesystem. A `find … | head -1` here
|
|
# once returned whichever Containerfile the directory listing happened to
|
|
# put first (dankclaw/alloy/Containerfile), so the test failed for sessions
|
|
# while the real ComfyUI Containerfile was correct, and it would equally
|
|
# have passed on a wrong file. Compose no longer names it: ComfyUI is a
|
|
# Quadlet unit and the image is built by deploy_comfyui.
|
|
local cf="$REPO_ROOT/dankclaw/Containerfile.comfyui"
|
|
grep -q 'containerfile=dankclaw/Containerfile.comfyui' "$REPO_ROOT/scripts/deploy-remote.sh" \
|
|
|| { echo "deployer no longer builds that Containerfile"; return 1; }
|
|
[ -f "$cf" ] || { echo "$cf does not exist"; return 1; }
|
|
grep -q "gfx1151" "$cf" || { echo "$cf does not reference gfx1151"; return 1; }
|
|
grep -q "FROM" "$cf" || { echo "$cf has no FROM"; return 1; }
|
|
}
|
|
|
|
@test "comfyui passes /dev/kfd and /dev/dri to the container" {
|
|
# Read the Quadlet unit, not compose: ComfyUI left compose on 2026-09-16
|
|
# after a stale, device-less container was restarted 9,146 times by
|
|
# `restart: unless-stopped` (see comfyui/quadlet/comfyui.container).
|
|
local unit="$REPO_ROOT/dankclaw/comfyui/quadlet/comfyui.container"
|
|
[ -f "$unit" ]
|
|
grep -q '^AddDevice=/dev/kfd$' "$unit" || { echo "no /dev/kfd"; return 1; }
|
|
grep -q '^AddDevice=/dev/dri$' "$unit" || { echo "no /dev/dri"; return 1; }
|
|
}
|
|
|
|
@test "comfyui does not run with --cpu" {
|
|
local unit="$REPO_ROOT/dankclaw/comfyui/quadlet/comfyui.container"
|
|
local exec_line
|
|
exec_line=$(grep -E '^Exec=' "$unit")
|
|
case "$exec_line" in
|
|
*--cpu*) echo "comfyui must not pass --cpu (ROCm GPU build)"; return 1 ;;
|
|
esac
|
|
case "$exec_line" in
|
|
*--listen*) ;;
|
|
*) echo "comfyui must listen on 0.0.0.0"; return 1 ;;
|
|
esac
|
|
}
|
|
|
|
@test "comfyui GPU access uses keep-groups + udev MODE 0666, not named groups" {
|
|
python3 - "$REPO_ROOT/dankclaw/comfyui/quadlet/comfyui.container" "$REPO_ROOT/dankclaw/udev/99-comfyui-gpu-access.rules" <<'PYEOF'
|
|
import sys, re
|
|
unit = open(sys.argv[1]).read()
|
|
groups = re.findall(r"^GroupAdd=(.*)$", unit, re.M)
|
|
# User-unit spawns carry NO supplementary groups, so named group_add entries
|
|
# (video, render) can never be honored at spawn time. The deployed contract is
|
|
# keep-groups plus world-writable compute nodes via the udev rule; a future
|
|
# edit that drops either half must fail here.
|
|
assert groups == ["keep-groups"], \
|
|
f"comfyui GroupAdd must be exactly ['keep-groups']; got: {groups}"
|
|
rule = open(sys.argv[2]).read()
|
|
assert 'KERNEL=="kfd", MODE="0666"' in rule, \
|
|
"udev rule must grant MODE 0666 on kfd with NO SUBSYSTEM guard (misc-class devices have no subsystem to match)"
|
|
assert 'KERNEL=="renderD*", SUBSYSTEM=="drm", MODE="0666"' in rule, \
|
|
"udev rule must grant MODE 0666 on renderD* (SUBSYSTEM drm guard)"
|
|
print("GPU access contract OK")
|
|
PYEOF
|
|
}
|