CI: lowercase image repository name; compose example too

Docker registry names must be lowercase, so PUBLIC/gpu-turnstile was
rejected by buildx. The workflow now lowercases gitea.repository.

.gitea/workflows/ci.yml

@@ -37,13 +37,19 @@ jobs:
             exit 1
           fi
 
+      - name: Compute lowercase image name
+        id: meta
+        run: |
+          REPO=git.rambossek.at/$(echo "${{ gitea.repository }}" | tr '[:upper:]' '[:lower:]')
+          echo "image=$REPO" >> "$GITHUB_OUTPUT"
+
       - uses: docker/setup-buildx-action@v3
 
       - uses: docker/login-action@v3
         with:
           registry: git.rambossek.at
           username: ${{ gitea.actor }}
-          password: ${{ secrets.GITEA_TOKEN }}
+          password: ${{ secrets.REGISTRY_TOKEN }}
 
       - uses: docker/build-push-action@v6
         with:
@@ -52,5 +58,5 @@ jobs:
           build-args: |
             VERSION=${{ gitea.ref_name }}
           tags: |
-            git.rambossek.at/${{ gitea.repository }}:${{ gitea.ref_name }}
-            git.rambossek.at/${{ gitea.repository }}:latest
+            ${{ steps.meta.outputs.image }}:${{ gitea.ref_name }}
+            ${{ steps.meta.outputs.image }}:latest

.gitignore

@@ -1,2 +1,4 @@
 /gpu-turnstile
 /gpu-turnstile.exe
+/compose.yaml
+/compose.yml

README.md

@@ -43,6 +43,13 @@ startup.
 | `WARM_MODEL` | _(empty)_ | Model to reload after an image job (off by default) |
 | `LOG_LEVEL` | `info` | `debug` logs every lock transition |
 | `LOG_FORMAT` | `text` | `json` for structured JSON logs |
+| `UNLOAD_POLL_INTERVAL` | `500ms` | `/api/ps` poll interval while unloading |
+| `HISTORY_POLL_INTERVAL` | `1s` | `/history/<id>` poll interval while a job runs |
+| `PROBE_TIMEOUT` | `5s` | Startup probe of both upstreams |
+| `FREE_TIMEOUT` | `30s` | `POST /free` call after an image job |
+| `WARM_TIMEOUT` | `2m` | Warm-model reload after an image job |
+| `SHUTDOWN_TIMEOUT` | `10s` | Graceful shutdown on SIGINT/SIGTERM |
+| `PROMPT_CAPTURE_LIMIT` | `65536` | Bytes of the `/prompt` response buffered to find `prompt_id` (pass-through is unaffected) |
 
 ## Observability
 
@@ -73,6 +80,10 @@ runs `go vet` and `go test -race`, and pushing a semantic-version tag
 `git.rambossek.at/<owner>/gpu-turnstile:vX.Y.Z` (and updates `:latest`).
 No images are built from branches.
 
+The registry login needs one repository secret (Settings → Actions →
+Secrets): `REGISTRY_TOKEN` — an access token with `write:package` scope.
+The automatic `GITEA_TOKEN` cannot push packages.
+
 ## Development
 
 ```sh

SPEC.md

@@ -81,12 +81,14 @@ ComfyUI listener (`:8188` → `COMFY_URL`):
 2. Unload Ollama: `GET /api/ps`; for each model `POST /api/generate
    {"model":M,"keep_alive":0}`; if that returns non-2xx (embedding-only
    models), `POST /api/embed {"model":M,"input":"x","keep_alive":0}`. Poll
-   `/api/ps` every 500 ms until empty or `UNLOAD_TIMEOUT`. On timeout: log and
+   `/api/ps` every `UNLOAD_POLL_INTERVAL` (default 500 ms) until empty or
+   `UNLOAD_TIMEOUT`. On timeout: log and
    continue (degrade, don't fail the user's request).
 3. Forward the original request body to ComfyUI `/prompt`, return status,
    headers and body to the caller unchanged, flush.
 4. If the response is 200 and contains `prompt_id`: in a goroutine, poll
-   `GET /history/<prompt_id>` every 1 s until the entry has
+   `GET /history/<prompt_id>` every `HISTORY_POLL_INTERVAL` (default 1 s)
+   until the entry has
    `status.completed == true`, `status.status_str == "error"`, or
    `JOB_TIMEOUT`. Then `POST /free {"unload_models":true,"free_memory":true}`.
    Then release the image lock.
@@ -111,6 +113,13 @@ load time. Off by default.
 | `LLM_WAIT_TIMEOUT` | `10m` | max time an LLM request waits for the lock before 503 |
 | `WARM_MODEL` | `` | optional model to reload after an image job |
 | `LOG_LEVEL` | `info` | `debug` logs every lock transition |
+| `UNLOAD_POLL_INTERVAL` | `500ms` | `/api/ps` poll interval while unloading |
+| `HISTORY_POLL_INTERVAL` | `1s` | `/history/<id>` poll interval while a job runs |
+| `PROBE_TIMEOUT` | `5s` | startup probe of both upstreams |
+| `FREE_TIMEOUT` | `30s` | `POST /free` call after an image job |
+| `WARM_TIMEOUT` | `2m` | warm-model reload after an image job |
+| `SHUTDOWN_TIMEOUT` | `10s` | graceful shutdown on SIGINT/SIGTERM |
+| `PROMPT_CAPTURE_LIMIT` | `65536` | bytes of the `/prompt` response buffered to find `prompt_id` (pass-through is unaffected) |
 
 Startup fails fast on unparsable values. Both upstreams are probed once at
 start (`/api/version`, `/system_stats`); failure is logged, not fatal.
@@ -185,8 +194,12 @@ are new.
      available in the runner image
   2. on a version tag only (`vX.Y.Z`, enforced): build the image with buildx
      and push it to the Gitea registry
-     `git.rambossek.at/<owner>/gpu-turnstile` tagged `:<tag>` and `:latest`,
-     using the workflow token (`${{ secrets.GITEA_TOKEN }}` / `gitea.actor`)
+     `git.rambossek.at/<owner>/gpu-turnstile` tagged `:<tag>` and `:latest`
+     (the repository path is lowercased in the workflow; Docker registry
+     names must be lowercase).
+     Login uses the repo secret `REGISTRY_TOKEN` (an access token with
+     `write:package` scope) because the automatic `GITEA_TOKEN` cannot push
+     packages; the username is just `gitea.actor`.
 - Release: a git tag `vX.Y.Z` produces the versioned image; the Open WebUI
   compose pins that tag. No images are built from branches.
 
@@ -194,7 +207,7 @@ are new.
 
 ```yaml
   gpu-turnstile:
-    image: git.rambossek.at/<owner>/gpu-turnstile:v0.1.0
+    image: git.rambossek.at/<owner>/gpu-turnstile:v0.1.0  # owner lowercased, e.g. "public"
     environment:
       OLLAMA_URL: http://<workstation-ip>:11435
       COMFY_URL:  http://<workstation-ip>:8189

cmd/gpu-turnstile/main.go

@@ -10,6 +10,7 @@ import (
 	"net/http"
 	"os"
 	"os/signal"
+	"strconv"
 	"strings"
 	"syscall"
 	"time"
@@ -32,6 +33,15 @@ type config struct {
 	unloadTimeout  time.Duration
 	jobTimeout     time.Duration
 	llmWaitTimeout time.Duration
+
+	unloadPollInterval  time.Duration
+	historyPollInterval time.Duration
+	probeTimeout        time.Duration
+	freeTimeout         time.Duration
+	warmTimeout         time.Duration
+	shutdownTimeout     time.Duration
+	promptCaptureLimit  int64
+
 	warmModel string
 	logLevel  slog.Level
 	logJSON   bool
@@ -59,6 +69,15 @@ func loadConfig(getenv func(string) string) (config, error) {
 		unloadTimeout:  time.Minute,
 		jobTimeout:     15 * time.Minute,
 		llmWaitTimeout: 10 * time.Minute,
+
+		unloadPollInterval:  500 * time.Millisecond,
+		historyPollInterval: time.Second,
+		probeTimeout:        5 * time.Second,
+		freeTimeout:         30 * time.Second,
+		warmTimeout:         2 * time.Minute,
+		shutdownTimeout:     10 * time.Second,
+		promptCaptureLimit:  64 * 1024,
+
 		logLevel: slog.LevelInfo,
 	}
 	for _, e := range []struct {
@@ -82,11 +101,24 @@ func loadConfig(getenv func(string) string) (config, error) {
 		{"UNLOAD_TIMEOUT", &cfg.unloadTimeout},
 		{"JOB_TIMEOUT", &cfg.jobTimeout},
 		{"LLM_WAIT_TIMEOUT", &cfg.llmWaitTimeout},
+		{"UNLOAD_POLL_INTERVAL", &cfg.unloadPollInterval},
+		{"HISTORY_POLL_INTERVAL", &cfg.historyPollInterval},
+		{"PROBE_TIMEOUT", &cfg.probeTimeout},
+		{"FREE_TIMEOUT", &cfg.freeTimeout},
+		{"WARM_TIMEOUT", &cfg.warmTimeout},
+		{"SHUTDOWN_TIMEOUT", &cfg.shutdownTimeout},
 	} {
 		if err := envDuration(getenv, e.name, e.dst); err != nil {
 			return cfg, err
 		}
 	}
+	if v := getenv("PROMPT_CAPTURE_LIMIT"); v != "" {
+		n, err := strconv.ParseInt(v, 10, 64)
+		if err != nil || n < 0 {
+			return cfg, fmt.Errorf("PROMPT_CAPTURE_LIMIT: must be a non-negative integer (bytes)")
+		}
+		cfg.promptCaptureLimit = n
+	}
 	if v := getenv("LOG_LEVEL"); v != "" {
 		var level slog.Level
 		if err := level.UnmarshalText([]byte(v)); err != nil {
@@ -149,6 +181,11 @@ func main() {
 		LLMWaitTimeout:      cfg.llmWaitTimeout,
 		UnloadTimeout:       cfg.unloadTimeout,
 		JobTimeout:          cfg.jobTimeout,
+		UnloadPollInterval:  cfg.unloadPollInterval,
+		HistoryPollInterval: cfg.historyPollInterval,
+		FreeTimeout:         cfg.freeTimeout,
+		WarmTimeout:         cfg.warmTimeout,
+		PromptCaptureLimit:  cfg.promptCaptureLimit,
 		WarmModel:           cfg.warmModel,
 	})
 	if err != nil {
@@ -157,7 +194,7 @@ func main() {
 	}
 
 	// Probe both upstreams once; failure is logged, not fatal.
-	probeCtx, probeCancel := context.WithTimeout(context.Background(), 5*time.Second)
+	probeCtx, probeCancel := context.WithTimeout(context.Background(), cfg.probeTimeout)
 	if err := ollamaClient.Probe(probeCtx); err != nil {
 		log.Warn("ollama probe failed", "url", cfg.ollamaURL, "err", err)
 	}
@@ -186,7 +223,7 @@ func main() {
 		log.Info("shutting down")
 	}
 
-	shutdownCtx, shutdownCancel := context.WithTimeout(context.Background(), 10*time.Second)
+	shutdownCtx, shutdownCancel := context.WithTimeout(context.Background(), cfg.shutdownTimeout)
 	defer shutdownCancel()
 	ollamaSrv.Shutdown(shutdownCtx)
 	comfySrv.Shutdown(shutdownCtx)

compose.yaml.example

@@ -0,0 +1,25 @@
+# Example deployment for gpu-turnstile. Copy to compose.yaml and adjust.
+#
+# gpu-turnstile listens on the ports the services normally use; the actual
+# Ollama and ComfyUI instances run one port higher on the workstation.
+services:
+  gpu-turnstile:
+    image: git.rambossek.at/public/gpu-turnstile:v0.1.1
+    restart: unless-stopped
+    environment:
+      OLLAMA_URL: http://<workstation-ip>:11435
+      COMFY_URL: http://<workstation-ip>:8189
+      # UNLOAD_TIMEOUT: 60s
+      # JOB_TIMEOUT: 15m
+      # LLM_WAIT_TIMEOUT: 10m
+      # WARM_MODEL: qwen3:14b
+      # LOG_LEVEL: info
+      # LOG_FORMAT: json
+    ports:
+      - "11434:11434" # LiteLLM api_base -> http://gpu-turnstile:11434
+      - "8188:8188"   # Open WebUI COMFYUI_BASE_URL -> http://gpu-turnstile:8188
+    networks: [internal]
+
+networks:
+  internal:
+    external: true

internal/proxy/proxy.go

@@ -21,10 +21,10 @@ import (
 	"gpu-turnstile/internal/ollama"
 )
 
-// captureLimit bounds how much of a /prompt response body is buffered while
-// looking for prompt_id. The body still passes through to the client
-// unchanged regardless of size.
-const captureLimit = 64 * 1024
+// defaultCaptureLimit bounds how much of a /prompt response body is
+// buffered while looking for prompt_id. The body still passes through to
+// the client unchanged regardless of size.
+const defaultCaptureLimit = 64 * 1024
 
 // Config wires a Server.
 type Config struct {
@@ -40,6 +40,18 @@ type Config struct {
 	LLMWaitTimeout time.Duration
 	UnloadTimeout  time.Duration
 	JobTimeout     time.Duration
+
+	// UnloadPollInterval and HistoryPollInterval override the clients'
+	// /api/ps and /history poll intervals when > 0.
+	UnloadPollInterval  time.Duration
+	HistoryPollInterval time.Duration
+	// FreeTimeout and WarmTimeout bound the /free call and the warm-model
+	// reload; zero selects the defaults.
+	FreeTimeout time.Duration
+	WarmTimeout time.Duration
+	// PromptCaptureLimit overrides defaultCaptureLimit when > 0.
+	PromptCaptureLimit int64
+
 	WarmModel string
 }
 
@@ -47,6 +59,9 @@ type Config struct {
 type Server struct {
 	cfg          Config
 	log          *slog.Logger
+	freeTimeout  time.Duration
+	warmTimeout  time.Duration
+	captureLimit int64
 
 	ollamaProxy *httputil.ReverseProxy
 	comfyProxy  *httputil.ReverseProxy
@@ -66,9 +81,30 @@ func New(cfg Config) (*Server, error) {
 	if log == nil {
 		log = slog.Default()
 	}
+	if cfg.UnloadPollInterval > 0 {
+		cfg.Ollama.PollInterval = cfg.UnloadPollInterval
+	}
+	if cfg.HistoryPollInterval > 0 {
+		cfg.Comfy.PollInterval = cfg.HistoryPollInterval
+	}
+	freeTimeout := cfg.FreeTimeout
+	if freeTimeout <= 0 {
+		freeTimeout = 30 * time.Second
+	}
+	warmTimeout := cfg.WarmTimeout
+	if warmTimeout <= 0 {
+		warmTimeout = 2 * time.Minute
+	}
+	captureLimit := cfg.PromptCaptureLimit
+	if captureLimit <= 0 {
+		captureLimit = defaultCaptureLimit
+	}
 	return &Server{
 		cfg:          cfg,
 		log:          log,
+		freeTimeout:  freeTimeout,
+		warmTimeout:  warmTimeout,
+		captureLimit: captureLimit,
 		ollamaProxy:  newReverseProxy(ollamaURL, log.With("upstream", "ollama")),
 		comfyProxy:   newReverseProxy(comfyURL, log.With("upstream", "comfy")),
 	}, nil
@@ -170,11 +206,12 @@ func (s *Server) ComfyHandler() http.Handler {
 }
 
 // captureWriter passes the response through unchanged while recording the
-// status code and the first captureLimit bytes of the body.
+// status code and the first limit bytes of the body.
 type captureWriter struct {
 	http.ResponseWriter
 	status int
 	buf    bytes.Buffer
+	limit  int64
 }
 
 func (w *captureWriter) WriteHeader(code int) {
@@ -183,7 +220,7 @@ func (w *captureWriter) WriteHeader(code int) {
 }
 
 func (w *captureWriter) Write(p []byte) (int, error) {
-	if w.buf.Len() < captureLimit {
+	if int64(w.buf.Len()) < w.limit {
 		w.buf.Write(p)
 	}
 	return w.ResponseWriter.Write(p)
@@ -228,7 +265,7 @@ func (s *Server) handlePrompt(w http.ResponseWriter, r *http.Request) {
 		log.Info("ollama models unloaded", "seconds", elapsed.Seconds())
 	}
 
-	cw := &captureWriter{ResponseWriter: w, status: http.StatusOK}
+	cw := &captureWriter{ResponseWriter: w, status: http.StatusOK, limit: s.captureLimit}
 	s.comfyProxy.ServeHTTP(cw, r)
 
 	var accepted struct {
@@ -262,7 +299,7 @@ func (s *Server) finishImageJob(promptID string) {
 		log.Info("image job completed")
 	}
 
-	freeCtx, freeCancel := context.WithTimeout(context.Background(), 30*time.Second)
+	freeCtx, freeCancel := context.WithTimeout(context.Background(), s.freeTimeout)
 	if err := s.cfg.Comfy.Free(freeCtx); err != nil {
 		log.Warn("failed to free ComfyUI models", "err", err)
 	}
@@ -273,7 +310,7 @@ func (s *Server) finishImageJob(promptID string) {
 
 	if s.cfg.WarmModel != "" {
 		if state, _, _ := s.cfg.Lock.Snapshot(); state == lock.StateIdle {
-			wctx, wcancel := context.WithTimeout(context.Background(), 2*time.Minute)
+			wctx, wcancel := context.WithTimeout(context.Background(), s.warmTimeout)
 			if err := s.cfg.Ollama.Warm(wctx, s.cfg.WarmModel); err != nil {
 				log.Warn("warm model reload failed", "model", s.cfg.WarmModel, "err", err)
 			} else {