9 Commits
Author SHA1 Message Date
mram 21e40a1774 Pin compose example to v0.2.0
ci / test (push) Successful in 14s
ci / docker (push) Successful in 1m7s
ci / release (push) Successful in 15s
2026-09-21 21:47:48 +02:00
mram bdc844872d Install narrates itself: stop/copy/start steps and the potentially long icacls tree grant 2026-09-21 21:36:45 +02:00
mram 299b6dc0bb Distinguish EACCES from ENOENT in the COMFY_DIR startup check 2026-09-21 21:32:45 +02:00
mram 9589e58ce6 Install opens up COMFY_DIR for the sandboxed service: ACL grant on Windows, BindPaths on Linux 2026-09-21 21:30:35 +02:00
mram be0bb36317 Pin compose example to v0.1.10
ci / test (push) Successful in 28s
ci / docker (push) Successful in 1m15s
ci / release (push) Successful in 17s
2026-09-21 21:20:49 +02:00
mram 9b267533c9 COMFY_DIR alone manages ComfyUI: derive the launch command from the standard venv layout 2026-09-21 21:18:33 +02:00
mram 6f092ddc12 Default GAME_POLL_INTERVAL to 15s: nvidia-smi polls keep the GPU awake 2026-09-21 21:10:21 +02:00
mram 0228ccc296 Pin compose example to v0.1.9
ci / test (push) Successful in 17s
ci / docker (push) Successful in 1m17s
ci / release (push) Successful in 19s
2026-09-21 20:57:34 +02:00
mram 9997913929 Sync the env file at startup, not just at install: updates append new settings 2026-09-21 20:56:32 +02:00
13 changed files with 332 additions and 68 deletions
+31 -12
View File
@@ -58,14 +58,14 @@ override file values. Invalid values fail at startup.
| `LLM_BUSY_STATUS` | `503` | HTTP status for rejected LLM requests in reject mode (400599, e.g. 429) |
| `BUSY_RETRY_AFTER` | `30` | Seconds sent as `Retry-After` on busy responses (both modes) |
| `WARM_MODEL` | _(empty)_ | Model to reload after an image job (off by default) |
| `COMFY_CMD` | _(empty = unmanaged)_ | Supervise ComfyUI: start on demand, stop when idle to free VRAM. Requires `COMFY_URL` |
| `COMFY_DIR` | _(empty)_ | Working directory for `COMFY_CMD` |
| `COMFY_CMD` | _(derived from `COMFY_DIR`; both empty = unmanaged)_ | Supervise ComfyUI: start on demand, stop when idle to free VRAM. Requires `COMFY_URL` |
| `COMFY_DIR` | _(empty)_ | Standard venv install root: set alone to supervise ComfyUI with the derived command (`.venv` + `main.py`); also the working directory for `COMFY_CMD` |
| `COMFY_IDLE_TIMEOUT` | `5m` | Stop the managed ComfyUI after this long idle |
| `COMFY_START_TIMEOUT` | `2m` | Max wait for the managed ComfyUI to come up |
| `GAME_PROCS` | _(empty = disabled)_ | Process names (comma-separated); while any runs, the GPU counts as held: requests wait, Ollama unloads, managed ComfyUI stops |
| `GPU_FOREIGN_VRAM_MB` | `0` (disabled) | Also treat the GPU as held when a non-ignored process uses more VRAM than this (needs nvidia-smi) |
| `GPU_IGNORE_PROCS` | `ollama,ollama app,ollama_llama_server,python,pythonw` | Process names never counted as foreign GPU users |
| `GAME_POLL_INTERVAL` | `5s` | How often game/VRAM detection runs |
| `GAME_POLL_INTERVAL` | `15s` | How often game/VRAM detection runs (don't go below ~10s — nvidia-smi polls keep the GPU awake) |
| `LOGLEVEL` | `warn` | `info` logs every request (colored arrows in text mode), `debug` adds lock transitions. `LOG_LEVEL` works as an alias |
| `LOG_FORMAT` | `text` | `json` for structured JSON logs |
| `LOG_FILE` | _(empty)_ | Append logs to this file instead of stderr |
@@ -99,18 +99,28 @@ override file values. Invalid values fail at startup.
class) in text mode, which renders in `docker compose logs` on Windows
Terminal. Set `NO_COLOR` to disable colors.
## Managed ComfyUI (`COMFY_CMD`)
## Managed ComfyUI (`COMFY_CMD` / `COMFY_DIR`)
Don't want ComfyUI running 24/7 (it holds VRAM even when idle — and the
Desktop app kills its server when you close it)? Point `COMFY_CMD` at a
standalone launch command and gpu-turnstile supervises it: the first
request starts it, it stops again after `COMFY_IDLE_TIMEOUT` (default 5m)
without work, freeing the GPU for games or the LLM. Example for a Desktop
install (run it once manually to confirm it works):
Desktop app kills its server when you close it)? gpu-turnstile can supervise
it: the first request starts it, it stops again after `COMFY_IDLE_TIMEOUT`
(default 5m) without work, freeing the GPU for games or the LLM.
The easy way — point `COMFY_DIR` at a standard venv install (a folder with
`.venv` and `main.py`, or `.venv` and `ComfyUI\main.py`) and the launch
command is derived from it, including `--port` from `COMFY_URL`:
```
COMFY_URL=http://127.0.0.1:8188
COMFY_CMD="C:\ComfyUI\.venv\Scripts\python.exe ComfyUI\main.py --port 8188"
COMFY_URL=http://127.0.0.1:8189
COMFY_DIR=C:\ComfyUI
```
For other layouts, spell the command out yourself (run it once manually to
confirm it works):
```
COMFY_URL=http://127.0.0.1:8189
COMFY_CMD="C:\ComfyUI\.venv\Scripts\python.exe" main.py --port 8189
COMFY_DIR=C:\ComfyUI
```
@@ -124,13 +134,22 @@ answers on the port, gpu-turnstile just uses it instead of spawning
instance already holds the port when you open the desktop app, the
desktop's server is the one that fails to bind.
One catch when gpu-turnstile runs as a service: the sandboxed service
account may not enter your user profile, so a ComfyUI install under
`C:\Users\...` (or `/home/...`) fails with "Access is denied".
`--install-service` fixes that automatically — it grants
`NT SERVICE\gpu-turnstile` recursive access to `COMFY_DIR` on Windows and
adds a `BindPaths=` to the systemd unit on Linux. Re-run it after changing
`COMFY_DIR`; or grant by hand from an admin shell:
`icacls "<COMFY_DIR>" /grant "NT SERVICE\gpu-turnstile:(OI)(CI)M" /T`.
## Game detection
Want to game on the same GPU without Ollama/ComfyUI squatting on the VRAM?
gpu-turnstile can watch for foreign GPU holders and, while one is active,
make LLM/image requests wait (or 503, per `LLM_BUSY_MODE`), unload Ollama's
models and stop the managed ComfyUI so the game gets the memory. Two
detection paths, each optional, polled every `GAME_POLL_INTERVAL` (5s):
detection paths, each optional, polled every `GAME_POLL_INTERVAL` (15s):
```
GAME_PROCS=cyberpunk2077.exe,bg3.exe # the reliable way on Windows
+20 -7
View File
@@ -128,10 +128,15 @@ state is `idle`, send `POST /api/generate {"model":WARM_MODEL,"keep_alive":-1}`
with empty prompt to reload the chat model so the next chat doesn't pay the
load time. Off by default.
### Managed ComfyUI (`COMFY_CMD`)
### Managed ComfyUI (`COMFY_CMD` / `COMFY_DIR`)
When `COMFY_CMD` is set, gpu-turnstile runs ComfyUI as a supervised child
process instead of expecting an always-on server:
process instead of expecting an always-on server. Setting only `COMFY_DIR`
enables the same management with the launch command derived from the
standard venv layout under it (`.venv\Scripts\python.exe` on Windows,
`.venv/bin/python` on Linux; `ComfyUI\main.py`, or a flat `main.py` when
that is what exists; `--port` from the `COMFY_URL` port). Missing layout
files are flagged in the startup log.
- **Start on demand**: any ComfyUI request spawns it (double quotes in the
command line group arguments with spaces; `COMFY_DIR` sets the working
@@ -154,12 +159,20 @@ process instead of expecting an always-on server:
- Its stdout/stderr is forwarded to the log at INFO. The health check
skips the intentionally-stopped/starting states; a failed probe while
the process is alive and was previously ready is logged as DOWN.
- **Permissions**: the service account is sandboxed (Windows virtual
account, systemd `DynamicUser`), so a ComfyUI install inside a user
profile is off-limits by default. `--install-service` opens it up —
a recursive ACL grant for `NT SERVICE\gpu-turnstile` on Windows, a
`BindPaths=` in the unit on Linux — reading `COMFY_DIR` from the config
it installs. Re-run `--install-service` after changing `COMFY_DIR`, or
grant by hand (admin shell):
`icacls "<COMFY_DIR>" /grant "NT SERVICE\gpu-turnstile:(OI)(CI)M" /T`.
## Game detection (foreign GPU holders)
Games and other foreign GPU users sit outside the URL-based consumer model —
nothing proxies through gpu-turnstile for them. Two independent detection
paths, polled every `GAME_POLL_INTERVAL` (default 5 s); either one being
paths, polled every `GAME_POLL_INTERVAL` (default 15 s); either one being
configured enables the feature:
- **Process watch list** (`GAME_PROCS`, comma-separated, case-insensitive,
@@ -211,14 +224,14 @@ override file values. A missing file is fine; a malformed one is fatal.
| `LLM_BUSY_STATUS` | `503` | HTTP status for rejected LLM requests in reject mode (400599, e.g. 429) |
| `BUSY_RETRY_AFTER` | `30` | seconds sent as `Retry-After` on busy responses (both modes) |
| `WARM_MODEL` | `` | optional model to reload after an image job |
| `COMFY_CMD` | _(empty = unmanaged)_ | spawn and supervise ComfyUI on demand: first request starts it, idle stop after `COMFY_IDLE_TIMEOUT` frees its VRAM. Requires `COMFY_URL` |
| `COMFY_DIR` | `` | working directory for `COMFY_CMD` |
| `COMFY_CMD` | _(derived from `COMFY_DIR`; both empty = unmanaged)_ | spawn and supervise ComfyUI on demand: first request starts it, idle stop after `COMFY_IDLE_TIMEOUT` frees its VRAM. Requires `COMFY_URL` |
| `COMFY_DIR` | `` | standard venv install root: set alone to supervise ComfyUI with the derived launch command (`.venv` + `main.py`, `--port` from `COMFY_URL`); also the working directory for `COMFY_CMD` |
| `COMFY_IDLE_TIMEOUT` | `5m` | stop the managed ComfyUI after this long without requests or jobs |
| `COMFY_START_TIMEOUT` | `2m` | how long a request waits for the managed ComfyUI to come up |
| `GAME_PROCS` | _(empty = disabled)_ | comma-separated process names (case-insensitive, `.exe` optional); while any runs, the GPU counts as held by it: requests wait, Ollama unloads, the managed ComfyUI stops |
| `GPU_FOREIGN_VRAM_MB` | `0` (disabled) | also treat the GPU as held when a process not in `GPU_IGNORE_PROCS` uses more VRAM than this; needs nvidia-smi |
| `GPU_IGNORE_PROCS` | `ollama,ollama app,ollama_llama_server,python,pythonw` | process names never counted as foreign GPU users |
| `GAME_POLL_INTERVAL` | `5s` | how often game/VRAM detection runs |
| `GAME_POLL_INTERVAL` | `15s` | how often game/VRAM detection runs (nvidia-smi polls keep the GPU awake; don't go below ~10s) |
| `LOGLEVEL` | `warn` | `info` logs every request (colored arrows in text mode), `debug` adds lock transitions. `LOG_LEVEL` is accepted as an alias |
| `LOG_FORMAT` | `text` | `json` for structured JSON logs |
| `LOG_FILE` | `` | append logs to this file instead of stderr (useful as a service) |
@@ -237,7 +250,7 @@ override file values. A missing file is fine; a malformed one is fatal.
| `UPDATE_REPO` | `https://git.rambossek.at/PUBLIC/gpu-turnstile` | repository to check for releases |
| `UPDATE_ASSET` | `gpu-turnstile.exe` | release asset to download |
| `APP_VER` | `stable` | version to run: `dev` disables updates, `stable` tracks the latest release, or an exact `vX.Y.Z` pin (up- or downgraded to) |
| `CFG_VER` | _(installer-managed)_ | config format reference written by `--install-service` (always a concrete `vX.Y.Z`; a dev build stamps `v0.0.0`); missing = the file is replaced with a fresh sample (backup `.bak`) |
| `CFG_VER` | _(installer-managed)_ | config format reference written by `--install-service` (always a concrete `vX.Y.Z`; a dev build stamps `v0.0.0`); missing = the file is replaced with a fresh sample (backup `.bak`). New settings are appended (commented out) at install and at every startup after an update changed the version |
Startup fails fast on unparsable values and when neither consumer URL is
set. Enabled upstreams are probed once at start (`/api/version`,
+62 -6
View File
@@ -15,6 +15,7 @@ import (
"os"
"os/signal"
"path/filepath"
"runtime"
"strings"
"syscall"
"time"
@@ -172,6 +173,7 @@ func main() {
}
log, logOut, logCloser := newLogger(cfg)
defer logCloser.Close()
syncEnvFile(resolveConfigPath(configPath), cfg.LogFile, log)
if service.IsService() {
if err := service.Run(func(ctx context.Context) error { return run(ctx, cfg, log, logOut, true) }); err != nil {
@@ -188,6 +190,27 @@ func main() {
}
}
// syncEnvFile upgrades an installer-written config file after an update:
// settings added since its CFG_VER are appended (commented out) and CFG_VER
// is bumped. Files not written by the installer (no CFG_VER), up-to-date
// files and dev builds are left untouched; a write failure is logged, not
// fatal.
func syncEnvFile(path, logFile string, log *slog.Logger) {
data, err := os.ReadFile(path)
if err != nil {
return // no config file; nothing to upgrade
}
synced, changed := config.SyncSample(string(data), version, logFile)
if !changed {
return
}
if err := os.WriteFile(path, []byte(synced), 0o644); err != nil {
log.Warn("could not append new settings to the config file", "path", path, "err", err)
return
}
log.Warn("config file updated: new settings appended", "path", path, "version", version)
}
// defaultConfigPath returns gpu-turnstile.env next to the executable.
func defaultConfigPath() string {
exe, err := os.Executable()
@@ -412,7 +435,26 @@ func orDisabled(url string) string {
return url
}
// managedComfyCommand resolves how ComfyUI is launched when it is managed:
// COMFY_CMD verbatim, or the standard venv layout under COMFY_DIR. Empty
// when neither is set (unmanaged).
func managedComfyCommand(cfg config.Config) string {
if cfg.ComfyCmd != "" {
return cfg.ComfyCmd
}
if cfg.ComfyDir != "" {
return supervise.DefaultComfyCommand(runtime.GOOS, cfg.ComfyDir, cfg.ComfyURL)
}
return ""
}
func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Writer, isService bool) error {
// ComfyUI can run as a managed child — COMFY_CMD verbatim, or the
// standard venv layout derived from COMFY_DIR alone: started on demand
// by the proxy, stopped after COMFY_IDLE_TIMEOUT idle (and on shutdown)
// so its VRAM is freed.
comfyCmdLine := managedComfyCommand(cfg)
// The startup line carries the version and every setting and is emitted
// at WARN so it is visible even with the default (quiet) log level.
log.Log(ctx, slog.LevelWarn, "starting gpu-turnstile",
@@ -438,7 +480,7 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
"backoff_max", cfg.BackoffMax,
"prompt_capture_limit", cfg.PromptCaptureLimit,
"warm_model", cfg.WarmModel,
"comfy_cmd", cfg.ComfyCmd,
"comfy_cmd", orDisabled(comfyCmdLine),
"comfy_dir", cfg.ComfyDir,
"comfy_idle_timeout", cfg.ComfyIdleTimeout,
"comfy_start_timeout", cfg.ComfyStartTimeout,
@@ -472,13 +514,27 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
}
}
// With COMFY_CMD set, ComfyUI runs as a managed child: started on
// demand by the proxy, stopped after COMFY_IDLE_TIMEOUT idle (and on
// shutdown) so its VRAM is freed.
var comfySup *supervise.Process
if cfg.ComfyCmd != "" {
if comfyCmdLine != "" {
if cfg.ComfyCmd == "" {
// Derived from COMFY_DIR: flag a wrong-looking layout early,
// while the operator is still watching the startup log. Inside
// a profile the service account may not enter, os.Stat fails
// with EACCES — that reads as "not found" but means "grant
// access", so say so.
python, script := supervise.ComfyLayout(runtime.GOOS, cfg.ComfyDir)
for _, p := range []string{python, script} {
if _, err := os.Stat(p); err != nil {
if isPermission(err) {
log.Warn("COMFY_DIR: not accessible to the service account; grant access or re-run --install-service", "path", p)
} else {
log.Warn("COMFY_DIR: file not found; ComfyUI requests will fail until it exists", "path", p)
}
}
}
}
var err error
comfySup, err = supervise.New("comfy", cfg.ComfyCmd, cfg.ComfyDir, comfyClient.Probe, cfg.ComfyStartTimeout, log)
comfySup, err = supervise.New("comfy", comfyCmdLine, cfg.ComfyDir, comfyClient.Probe, cfg.ComfyStartTimeout, log)
if err != nil {
return err
}
+1 -1
View File
@@ -6,7 +6,7 @@
# ComfyUI --listen 0.0.0.0 --port 8189).
services:
gpu-turnstile:
image: git.rambossek.at/public/gpu-turnstile:v0.1.8
image: git.rambossek.at/public/gpu-turnstile:v0.2.0
restart: unless-stopped
environment:
# Each consumer is enabled by setting its URL; leave one unset to
+12 -6
View File
@@ -51,11 +51,14 @@ type Config struct {
LLMBusyStatus int
BusyRetryAfter int
// ComfyCmd spawns and supervises a ComfyUI server on demand (empty =
// unmanaged, the current behavior). ComfyDir is its working directory.
// The managed server is stopped after ComfyIdleTimeout without
// requests, freeing its VRAM; ComfyStartTimeout bounds how long a
// request waits for it to come up.
// ComfyCmd spawns and supervises a ComfyUI server on demand. When
// ComfyCmd is empty but ComfyDir is set, management is enabled with the
// standard venv layout under ComfyDir (.venv + main.py or
// ComfyUI/main.py; --port from the COMFY_URL port) — ComfyCmd is the
// override for other layouts and doubles as the working directory when
// set explicitly. The managed server is stopped after ComfyIdleTimeout
// without requests, freeing its VRAM; ComfyStartTimeout bounds how long
// a request waits for it to come up.
ComfyCmd string
ComfyDir string
ComfyIdleTimeout time.Duration
@@ -116,7 +119,7 @@ func Defaults() Config {
// ComfyUI runs under python; excluding it (and Ollama) by name keeps
// our own consumers from tripping the foreign-VRAM check.
GPUIgnoreProcs: []string{"ollama", "ollama app", "ollama_llama_server", "python", "pythonw"},
GamePollInterval: 5 * time.Second,
GamePollInterval: 15 * time.Second,
LogLevel: slog.LevelWarn,
}
@@ -305,6 +308,9 @@ func Load(getenv func(string) string) (Config, error) {
if cfg.ComfyCmd != "" && cfg.ComfyURL == "" {
return cfg, fmt.Errorf("COMFY_CMD requires COMFY_URL to be set (the proxy needs somewhere to forward)")
}
if cfg.ComfyCmd == "" && cfg.ComfyDir != "" && cfg.ComfyURL == "" {
return cfg, fmt.Errorf("COMFY_DIR without COMFY_CMD requires COMFY_URL to be set (it enables the managed ComfyUI)")
}
if cfg.OllamaURL == "" && cfg.ComfyURL == "" {
return cfg, ErrNoConsumer
}
+27 -2
View File
@@ -201,8 +201,8 @@ func TestGameDetectionSettings(t *testing.T) {
if len(cfg.GPUIgnoreProcs) != 2 || cfg.GPUIgnoreProcs[1] != "my-trainer" {
t.Fatalf("GPUIgnoreProcs = %v", cfg.GPUIgnoreProcs)
}
if cfg.GamePollInterval != 5*time.Second {
t.Fatalf("GamePollInterval = %v, want 5s default", cfg.GamePollInterval)
if cfg.GamePollInterval != 15*time.Second {
t.Fatalf("GamePollInterval = %v, want 15s default", cfg.GamePollInterval)
}
// Defaults: both detection paths off, ignore list covers our consumers.
@@ -222,3 +222,28 @@ func TestGameDetectionSettings(t *testing.T) {
t.Fatal("GPUIgnoreProcs default must not be empty")
}
}
func TestComfyDirOnlyEnablesManaged(t *testing.T) {
// COMFY_DIR without COMFY_CMD and without COMFY_URL is a mistake.
_, err := Load(func(k string) string {
if k == "COMFY_DIR" {
return `C:\ComfyUI`
}
return ""
})
if err == nil || !strings.Contains(err.Error(), "COMFY_DIR") {
t.Fatalf("err = %v, want COMFY_DIR/COMFY_URL validation error", err)
}
// With COMFY_URL it loads — the launch command is derived from the dir.
if _, err := Load(func(k string) string {
switch k {
case "COMFY_DIR":
return `C:\ComfyUI`
case "COMFY_URL":
return "http://127.0.0.1:8189"
}
return ""
}); err != nil {
t.Fatalf("COMFY_DIR with COMFY_URL must load: %v", err)
}
}
+5 -4
View File
@@ -28,14 +28,14 @@ func sampleEntries(logFile string) []sampleEntry {
{"OLLAMA_URL", "http://127.0.0.1:11434", "Ollama upstream URL; setting it enables the Ollama consumer (default: empty = disabled)", false},
{"COMFY_URL", "http://127.0.0.1:8188", "ComfyUI upstream URL; setting it enables the ComfyUI consumer (default: empty = disabled)", false},
{"WARM_MODEL", "", "Optional model to reload after an image job (default: empty = none)", false},
{"COMFY_CMD", `"C:\ComfyUI\.venv\Scripts\python.exe" ComfyUI\main.py --port 8188`, "Spawn and supervise ComfyUI on demand: the first request starts it, it stops after COMFY_IDLE_TIMEOUT to free VRAM (default: empty = unmanaged)", false},
{"COMFY_DIR", `C:\ComfyUI`, "Working directory for COMFY_CMD (default: empty = inherit)", false},
{"COMFY_CMD", `"C:\ComfyUI\.venv\Scripts\python.exe" ComfyUI\main.py --port 8188`, "Spawn and supervise ComfyUI on demand: the first request starts it, it stops after COMFY_IDLE_TIMEOUT to free VRAM (default: derived from COMFY_DIR; both empty = unmanaged)", false},
{"COMFY_DIR", `C:\ComfyUI`, "Root of a standard ComfyUI venv install (.venv + main.py): setting it alone supervises ComfyUI with the derived launch command; also the working directory for COMFY_CMD", false},
{"COMFY_IDLE_TIMEOUT", "5m", "Stop the managed ComfyUI after this long without requests or jobs (frees VRAM)", false},
{"COMFY_START_TIMEOUT", "2m", "How long a request waits for the managed ComfyUI to come up", false},
{"GAME_PROCS", "cyberpunk2077.exe,hl2.exe", "While a listed process runs, the GPU counts as held by it: requests wait, Ollama unloads, managed ComfyUI stops (default: empty = disabled)", false},
{"GPU_FOREIGN_VRAM_MB", "1024", "Also treat the GPU as held when a process not in GPU_IGNORE_PROCS uses more VRAM than this (needs nvidia-smi; 0/empty = disabled)", false},
{"GPU_IGNORE_PROCS", "ollama,ollama app,ollama_llama_server,python,pythonw", "Process names never counted as foreign GPU users (ComfyUI runs under python)", false},
{"GAME_POLL_INTERVAL", "5s", "How often game/VRAM detection runs", false},
{"GAME_POLL_INTERVAL", "15s", "How often game/VRAM detection runs (nvidia-smi polls keep the GPU awake; don't go below ~10s)", false},
{"UNLOAD_TIMEOUT", "60s", "How long to wait for Ollama to unload a model", false},
{"JOB_TIMEOUT", "15m", "Maximum time to wait for a ComfyUI job", false},
{"LLM_WAIT_TIMEOUT", "10m", "Max time an LLM request waits for the GPU before being answered 503 (wait mode)", false},
@@ -66,7 +66,8 @@ func sampleEntries(logFile string) []sampleEntry {
// comment line. Everything is commented out — so all defaults apply —
// except the CFG_VER/APP_VER header and LOG_FILE when logFile is non-empty
// (a Windows service has no console). CFG_VER records the version that
// wrote the file so later installs can upgrade it.
// wrote the file so installs — and startups after an update — can upgrade
// it.
func SampleEnv(version, logFile string) string {
// CFG_VER is always a concrete vX.Y.Z — never "dev". A dev build
// stamps v0.0.0, which sorts older than any release, so the next
+16
View File
@@ -45,3 +45,19 @@ func writeEnvFile(path, content string) error {
}
return nil
}
// configuredValue reads one key from the env file the service will load, so
// the installers can adapt the sandbox to it (ACL grants, unit directives).
// "" when unset or unreadable.
func configuredValue(configPath, key string) string {
f, err := os.Open(configPath)
if err != nil {
return ""
}
defer f.Close()
values, err := config.ParseEnvFile(f)
if err != nil {
return ""
}
return values[key]
}
+14 -6
View File
@@ -62,8 +62,14 @@ func Run(run func(ctx context.Context) error) error {
// filesystem is read-only except StateDirectory (the install dir, so
// self-updates can rewrite the binary), and the usual no-privilege-escalation
// directives apply. The proxy needs nothing but outbound TCP/UDP and the
// notify socket, so it loses nothing.
func renderUnit(exePath, configPath string) string {
// notify socket, so it loses nothing. A managed ComfyUI (comfyDir) gets a
// BindPaths hole through ProtectHome/ProtectSystem: it reads its venv and
// writes output/temp/user data under COMFY_DIR.
func renderUnit(exePath, configPath, comfyDir string) string {
bind := ""
if comfyDir != "" {
bind = "BindPaths=" + comfyDir + "\n"
}
return fmt.Sprintf(`[Unit]
Description=gpu-turnstile GPU arbitration proxy for Ollama and ComfyUI
After=network-online.target
@@ -78,7 +84,7 @@ RestartSec=5s
DynamicUser=yes
StateDirectory=%s
ProtectSystem=strict
%sProtectSystem=strict
ProtectHome=yes
PrivateTmp=yes
NoNewPrivileges=yes
@@ -100,7 +106,7 @@ SystemCallErrorNumber=EPERM
[Install]
WantedBy=multi-user.target
`, exePath, configPath, Name)
`, exePath, configPath, Name, bind)
}
// copyFile copies src to dst, creating dst with the given mode.
@@ -125,7 +131,9 @@ func copyFile(src, dst string, mode os.FileMode) error {
// sure /etc/gpu-turnstile.env exists (copied from the given config file if
// provided), writes the hardened unit, then enables and starts it. With
// copyBin=false the current executable location and config path are
// registered as-is instead. Needs root.
// registered as-is instead. When the config sets COMFY_DIR, the unit gets a
// BindPaths= for it so the sandboxed service can reach the managed ComfyUI
// even under /home. Needs root.
//
// Re-running install converges an existing unit instead of failing: it is
// stopped first if active, the installed binary is replaced only when the
@@ -186,7 +194,7 @@ func Install(configPath string, copyBin bool, version string) error {
cfg = abs
}
}
rendered := renderUnit(exe, cfg)
rendered := renderUnit(exe, cfg, configuredValue(cfg, "COMFY_DIR"))
if old, _ := os.ReadFile(unitPath); string(old) != rendered {
if err := os.WriteFile(unitPath, []byte(rendered), 0o644); err != nil {
return fmt.Errorf("write %s (run as root): %w", unitPath, err)
+9 -1
View File
@@ -8,7 +8,7 @@ import (
)
func TestRenderUnit(t *testing.T) {
unit := renderUnit("/var/lib/gpu-turnstile/gpu-turnstile", "/etc/gpu-turnstile.env")
unit := renderUnit("/var/lib/gpu-turnstile/gpu-turnstile", "/etc/gpu-turnstile.env", "")
for _, want := range []string{
"Type=notify",
"WatchdogSec=30s",
@@ -26,4 +26,12 @@ func TestRenderUnit(t *testing.T) {
t.Fatalf("unit missing %q:\n%s", want, unit)
}
}
if strings.Contains(unit, "BindPaths") {
t.Fatalf("unit without COMFY_DIR must not bind anything:\n%s", unit)
}
unit = renderUnit("/var/lib/gpu-turnstile/gpu-turnstile", "/etc/gpu-turnstile.env", "/home/gpu/ComfyUI")
if !strings.Contains(unit, "BindPaths=/home/gpu/ComfyUI\n") {
t.Fatalf("unit with COMFY_DIR must bind it:\n%s", unit)
}
}
+45 -23
View File
@@ -24,8 +24,6 @@ import (
"golang.org/x/sys/windows"
"golang.org/x/sys/windows/svc"
"golang.org/x/sys/windows/svc/mgr"
"gpu-turnstile/internal/config"
)
// Name is the Windows service name.
@@ -106,10 +104,12 @@ func (h *handler) Execute(_ []string, requests <-chan svc.ChangeRequest, status
// the service after 5s on failure — this is also what brings up a staged
// update after the updater exits with a non-zero code. After registering,
// the virtual account is granted modify access to the install and data
// directories (self-updates rewrite the exe), and read access to the
// config file if it lives elsewhere. The grants must come after
// CreateService: the virtual account's SID only exists once the service is
// registered.
// directories (self-updates rewrite the exe), read access to the
// config file if it lives elsewhere, and — when the config sets COMFY_DIR —
// recursive modify access to the managed ComfyUI's install tree, which may
// live inside a user profile the account otherwise cannot enter. The grants
// must come after CreateService: the virtual account's SID only exists once
// the service is registered.
//
// Re-running install on an existing service converges instead of failing:
// the service is stopped first if running (so the binary can be replaced),
@@ -146,6 +146,7 @@ func Install(configPath string, copyBin bool, version string) error {
if st, qErr := s.Query(); qErr == nil &&
(st.State == svc.Running || st.State == svc.StartPending) {
wasRunning = true
fmt.Println("stopping the running gpu-turnstile service")
if err := stopAndWait(s); err != nil {
return err
}
@@ -161,6 +162,7 @@ func Install(configPath string, copyBin bool, version string) error {
}
installedExe := filepath.Join(installDir, "gpu-turnstile.exe")
if same, _ := sameFileContent(exe, installedExe); !same {
fmt.Printf("installing %s\n", installedExe)
if err := copyFile(exe, installedExe); err != nil {
return fmt.Errorf("copy binary to %s: %w", installedExe, err)
}
@@ -208,6 +210,7 @@ func Install(configPath string, copyBin bool, version string) error {
// Best effort: start now instead of waiting for the next boot. A
// missing config (no consumer URLs) fails the start; the service stays
// registered and can be started once the config exists.
fmt.Println("starting the gpu-turnstile service")
s.Start()
return nil
}
@@ -243,6 +246,7 @@ func Install(configPath string, copyBin bool, version string) error {
return err
}
if wasRunning {
fmt.Println("starting the gpu-turnstile service")
if err := s.Start(); err != nil {
return fmt.Errorf("start service: %w", err)
}
@@ -354,7 +358,7 @@ func grantAll(exe, configPath string) error {
return err
}
}
if logFile := configuredLogFile(configPath); logFile != "" {
if logFile := configuredValue(configPath, "LOG_FILE"); logFile != "" {
dir := filepath.Dir(logFile)
if err := os.MkdirAll(dir, 0o755); err == nil {
if err := grantAccess(dir, "(OI)(CI)(M)"); err != nil {
@@ -362,6 +366,17 @@ func grantAll(exe, configPath string) error {
}
}
}
// A managed ComfyUI whose install lives somewhere the virtual account
// may not go (a user profile) needs an explicit grant — recursively,
// since ComfyUI also writes output/temp/user data next to its code. A
// missing directory is skipped: the startup warning covers it.
if comfyDir := configuredValue(configPath, "COMFY_DIR"); comfyDir != "" {
if _, err := os.Stat(comfyDir); err == nil {
if err := grantAccessTree(comfyDir, "(OI)(CI)(M)"); err != nil {
return err
}
}
}
return nil
}
@@ -466,28 +481,35 @@ func RelaunchElevated(args []string) (int, error) {
// grantAccess gives the virtual account the icacls permission set (e.g.
// "(OI)(CI)(M)") on path.
func grantAccess(path, perms string) error {
out, err := exec.Command("icacls", path, "/grant", virtualAccount+":"+perms).CombinedOutput()
return runIcacls(path, perms, false)
}
// grantAccessTree is grantAccess with /T: the ACE is applied to the
// existing tree, not just inherited by children created later. Needed when
// the tree already exists, e.g. a ComfyUI install in a user profile. On a
// large tree (a venv has tens of thousands of files) this takes minutes,
// so it says what it is doing instead of looking hung.
func grantAccessTree(path, perms string) error {
return runIcacls(path, perms, true)
}
func runIcacls(path, perms string, recursive bool) error {
args := []string{path, "/grant", virtualAccount + ":" + perms}
if recursive {
fmt.Printf("granting %s modify access to %s (large trees can take minutes)\n", virtualAccount, path)
args = append(args, "/T")
}
start := time.Now()
out, err := exec.Command("icacls", args...).CombinedOutput()
if err != nil {
return fmt.Errorf("grant %s access to %s: %w (%s)", virtualAccount, path, err, strings.TrimSpace(string(out)))
}
if recursive {
fmt.Printf("access granted in %s\n", time.Since(start).Round(time.Second))
}
return nil
}
// configuredLogFile reads LOG_FILE from the config file so the installer
// can pre-create and ACL the log directory. "" when unset or unreadable.
func configuredLogFile(configPath string) string {
f, err := os.Open(configPath)
if err != nil {
return ""
}
defer f.Close()
values, err := config.ParseEnvFile(f)
if err != nil {
return ""
}
return values["LOG_FILE"]
}
// RestartIfRunning restarts the service when it is installed and running
// (used after a forced update staged a new binary). Reports whether a
// restart happened. A service that is not installed or not running is not
+41
View File
@@ -9,13 +9,54 @@ import (
"fmt"
"io"
"log/slog"
"net/url"
"os"
"os/exec"
"path/filepath"
"runtime"
"strings"
"sync"
"time"
)
// ComfyLayout returns the interpreter and script path of a standard ComfyUI
// venv install rooted at dir for the given GOOS: .venv\Scripts\python.exe
// on Windows, .venv/bin/python elsewhere. The script is main.py — either in
// a ComfyUI subdirectory or directly under dir, whichever exists (the
// subdirectory form wins ties and is the default when neither exists yet,
// so the caller's missing-file warning points at the documented layout).
func ComfyLayout(goos, dir string) (python, script string) {
if goos == "windows" {
python = filepath.Join(dir, ".venv", "Scripts", "python.exe")
} else {
python = filepath.Join(dir, ".venv", "bin", "python")
}
script = filepath.Join(dir, "ComfyUI", "main.py")
if _, err := os.Stat(script); err != nil {
if _, err := os.Stat(filepath.Join(dir, "main.py")); err == nil {
script = filepath.Join(dir, "main.py")
}
}
return python, script
}
// DefaultComfyCommand builds the launch command for the standard venv
// layout (see ComfyLayout): the script is passed relative to dir so dir
// stays the working directory, and --port is taken from comfyURL when the
// URL carries one.
func DefaultComfyCommand(goos, dir, comfyURL string) string {
python, script := ComfyLayout(goos, dir)
rel, err := filepath.Rel(dir, script)
if err != nil {
rel = script
}
cmd := `"` + python + `" ` + rel
if u, err := url.Parse(comfyURL); err == nil && u.Port() != "" {
cmd += " --port " + u.Port()
}
return cmd
}
// Process is one managed child process.
type Process struct {
name string
+49
View File
@@ -7,6 +7,8 @@ import (
"net/http"
"net/http/httptest"
"os"
"path/filepath"
"strings"
"testing"
"time"
)
@@ -190,3 +192,50 @@ func TestWatchIdleRespectsBusyGPU(t *testing.T) {
t.Fatal("process was stopped while the GPU was busy")
}
}
func TestComfyLayoutAndDefaultCommand(t *testing.T) {
// Nested layout (ComfyUI/main.py under dir) wins.
dir := t.TempDir()
nested := filepath.Join(dir, "ComfyUI", "main.py")
if err := os.MkdirAll(filepath.Dir(nested), 0o755); err != nil {
t.Fatal(err)
}
if err := os.WriteFile(nested, []byte("x"), 0o644); err != nil {
t.Fatal(err)
}
python, script := ComfyLayout("windows", dir)
if want := filepath.Join(dir, ".venv", "Scripts", "python.exe"); python != want {
t.Errorf("python = %s, want %s", python, want)
}
if script != nested {
t.Errorf("script = %s, want %s", script, nested)
}
cmd := DefaultComfyCommand("windows", dir, "http://127.0.0.1:8189")
want := `"` + filepath.Join(dir, ".venv", "Scripts", "python.exe") + `" ` + filepath.Join("ComfyUI", "main.py") + " --port 8189"
if cmd != want {
t.Errorf("cmd = %q, want %q", cmd, want)
}
// Flat layout (main.py directly under dir) is found too.
flat := t.TempDir()
if err := os.WriteFile(filepath.Join(flat, "main.py"), []byte("x"), 0o644); err != nil {
t.Fatal(err)
}
if _, script := ComfyLayout("linux", flat); script != filepath.Join(flat, "main.py") {
t.Errorf("flat script = %s", script)
}
cmd = DefaultComfyCommand("linux", flat, "http://comfy.internal")
if strings.Contains(cmd, "--port") {
t.Errorf("cmd = %q, want no --port for a port-less URL", cmd)
}
if !strings.HasSuffix(cmd, `" main.py`) {
t.Errorf("cmd = %q, want quoted python + relative main.py", cmd)
}
// Neither exists yet: default to the documented nested form so the
// startup warning points there.
empty := t.TempDir()
if _, script := ComfyLayout("windows", empty); script != filepath.Join(empty, "ComfyUI", "main.py") {
t.Errorf("missing-layout script = %s", script)
}
}