Prepare isolated Qwen Image 2.1 evaluation

2026-09-21 08:52:36 +02:00
parent 302b08e051
commit 1a660c20ba
8 changed files with 351 additions and 1 deletions
@@ -22,6 +22,8 @@ dokumentieren den ausgerollten Stand, reduzierte Statuslatenzen und Tests.
 - genau ein aktives llama.cpp-Profil: Fast, Medium, Large, Ultra oder Uncensored
 - Profile Router auf Port 8081
 - FLUX.2 Klein 9B FP8 Beta: Transformer/VAE auf RTX 5080, Textencoder auf RTX 3060
+- Qwen-Image-2.1 als isolierter, standardmäßig gestoppter Vergleichsworker
+  ([Testablauf](docs/QWEN_IMAGE_21_TEST.md)); nicht produktiv an den Router angebunden
 - Qwen3-TTS 1.7B auf der RTX 3060 hinter dem TTS-Gateway; kein Piper-Fallback
 - Whisper.cpp `small` auf der CPU für lokale deutsche Spracherkennung
 - EmbeddingGemma 300M Q8 auf der CPU für OpenClaws hybride Memory-Suche
@@ -572,6 +572,7 @@ services:
      ALLOWED_PROFILES: fast,medium,large,ultra,uncensored
      IMAGE_WORKER: image
      RESTORE_WORKER: restore
+      QWEN_IMAGE_TEST_WORKER: qwen-image-2.1-test
      TTS_WORKER: qwen3
      MUSIC_WORKER: acestep
      YUE2_WORKER: yue2
@@ -708,6 +709,69 @@ services:
      timeout: 3s
      retries: 12

+  # Isolated evaluation target. It is created in a stopped state by the
+  # preparation script and can only be started through the controller's
+  # allowlisted, GPU-exclusive test endpoint.
+  qwen-image-21-test:
+    build:
+      context: platform/docker/qwen-image-21-test
+      args:
+        COMFYUI_COMMIT: b0f4b7b294ce482a2e071d9d762c133d38c7aa07
+    image: mike-ai/qwen-image-2.1-test:local
+    container_name: mike-ai-qwen-image-2.1-test
+    restart: "no"
+    profiles: [qwen-image-test]
+    labels:
+      com.mike-ai.image-worker: qwen-image-2.1-test
+    gpus: all
+    read_only: true
+    tmpfs: ["/tmp:size=4g,mode=1777"]
+    volumes:
+      - "${QWEN_IMAGE_21_MODEL_DIR:-/data/models/Qwen-Image-2.1-ComfyUI}:/opt/ComfyUI/models:ro"
+      - "${QWEN_IMAGE_21_OUTPUT_DIR:-/data/qwen-image-2.1-test-output}:/opt/ComfyUI/output"
+    environment:
+      # Athena enumerates RTX 3060 as GPU 0 and RTX 5080 as GPU 1.
+      NVIDIA_VISIBLE_DEVICES: "${QWEN_IMAGE_21_GPU:-1}"
+      NVIDIA_DRIVER_CAPABILITIES: compute,utility
+      PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    command:
+      - python
+      - main.py
+      - --listen
+      - 0.0.0.0
+      - --port
+      - "8188"
+      - --lowvram
+      - --preview-method
+      - none
+    networks: [inference]
+    security_opt: ["no-new-privileges:true"]
+    cap_drop: [ALL]
+    healthcheck:
+      test: [CMD, python, -c, "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8188/system_stats', timeout=2)"]
+      interval: 5s
+      timeout: 3s
+      retries: 60
+      start_period: 10s
+
+  qwen-image-21-test-runner:
+    image: python:3.12-slim
+    profiles: [qwen-image-test]
+    read_only: true
+    tmpfs: ["/tmp:size=32m,mode=1777"]
+    volumes:
+      - ./scripts/qwen-image-21-test.py:/opt/test/run.py:ro
+      - "${QWEN_IMAGE_21_OUTPUT_DIR:-/data/qwen-image-2.1-test-output}:/output"
+    environment:
+      CONTROLLER_URL: http://profile-controller:8090
+      CONTROLLER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
+      COMFY_URL: http://qwen-image-21-test:8188
+      OUTPUT_DIR: /output
+    command: [python, /opt/test/run.py]
+    networks: [control, inference]
+    security_opt: ["no-new-privileges:true"]
+    cap_drop: [ALL]
+
  qwen3-tts:
    image: ${QWEN3_TTS_IMAGE:-ghcr.io/malaiwah/qwen3-tts-server:latest@sha256:b363a01d08b1bbecbfc3ca6f585368fae2cfdc591f9ecca6643738369f9a9d98}
    container_name: mike-ai-qwen3-tts
@@ -32,6 +32,11 @@ def restore_item(state="exited"):
            "Labels": {controller.IMAGE_LABEL_KEY: controller.RESTORE_WORKER}}


+def qwen_image_test_item(state="exited"):
+    return {"Id": "id-qwen-image-test", "State": state,
+            "Labels": {controller.IMAGE_LABEL_KEY: controller.QWEN_IMAGE_TEST_WORKER}}
+
+
 def tts_item(state="running"):
    return {"Id": "id-tts", "State": state,
            "Labels": {controller.TTS_LABEL_KEY: controller.TTS_WORKER}}
@@ -177,6 +182,33 @@ class ProfileControllerTests(unittest.TestCase):
            ("POST", "/containers/id-restore/start"),
        ])

+    def test_qwen_image_test_is_allowlisted_and_exclusive(self):
+        profiles = {name: item(name) for name in controller.ALLOWED}
+        profiles["medium"] = item("medium", "running")
+        calls = []
+
+        def request(method, path):
+            calls.append((method, path))
+            return 204, b""
+
+        with patch.object(controller, "containers", return_value=profiles), \
+             patch.object(controller, "image_container", return_value=qwen_image_test_item()), \
+             patch.object(controller, "image_containers", return_value=[
+                 image_item("running"), restore_item(), qwen_image_test_item()]), \
+             patch.object(controller, "tts_container", return_value=tts_item()), \
+             patch.object(controller, "docker_request", side_effect=request):
+            result = controller.set_image_worker(
+                True, controller.QWEN_IMAGE_TEST_WORKER)
+
+        self.assertEqual(result, {
+            "image_worker": "qwen-image-2.1-test", "state": "running"})
+        self.assertEqual(calls, [
+            ("POST", "/containers/id-medium/stop?t=120"),
+            ("POST", "/containers/id-tts/stop?t=30"),
+            ("POST", "/containers/id-flux/stop?t=20"),
+            ("POST", "/containers/id-qwen-image-test/start"),
+        ])
+
    def test_profile_activation_stops_image_worker_first(self):
        profiles = {name: item(name) for name in controller.ALLOWED}
        calls = []
@@ -0,0 +1,60 @@
+# Qwen-Image-2.1 – isolierter Test
+
+Stand: 21. September 2026. Dieser Pfad dient nur dem Vergleich mit dem
+produktiven FLUX.2-Worker. Er ersetzt FLUX nicht und ist nicht an den Router
+angebunden.
+
+## Reproduzierbarer Stand
+
+- ComfyUI: Commit `b0f4b7b294ce482a2e071d9d762c133d38c7aa07`
+- Modell-Repository: `Comfy-Org/Qwen-Image-2.1`, Revision
+  `ace0edeb3791a594ddfa36ed5f41a178a394e921`
+- Transformer: `qwen_image_2.1_int8_convrot.safetensors` (7.256.783.064 Byte)
+- Textencoder: `qwen3vl_8b_int8_convrot.safetensors` (9.350.798.360 Byte)
+- VAE: `qwen_image_2.1_vae_bf16.safetensors` (675.509.688 Byte)
+- erster Vergleich: 1024 × 1024, 25 Schritte, CFG 1, Euler/Simple
+
+Die Gewichte sind die von Qwen verlinkte offizielle ComfyUI-Aufbereitung. Der
+Worker verwendet `--lowvram` auf der RTX 5080. So bleibt der Test mit 16 GB
+VRAM möglich; der Preis ist CPU-Offload und damit eine längere Laufzeit.
+
+## Vorbereitung ohne GPU-Last
+
+```bash
+cd /opt/mike-ai/stack
+sudo ./scripts/prepare-qwen-image-21-test.sh
+```
+
+Das Skript prüft feste SHA-256-Summen, baut das Image, aktualisiert den
+Profile Controller und erstellt den Testcontainer **gestoppt**. Es lädt kein
+Modell in die GPU.
+
+## Einmaliger Test nach ausdrücklichem „Go“
+
+```bash
+cd /opt/mike-ai/stack
+docker compose --env-file /etc/mike-ai/stack.env \
+  --profile qwen-image-test run --rm qwen-image-21-test-runner \
+  python /opt/test/run.py 'HIER DEN VEREINBARTEN PROMPT EINSETZEN'
+```
+
+Der Runner merkt sich das aktive LLM-Profil, startet den Qwen-Worker über den
+allowlist-beschränkten Controller, erzeugt genau ein Bild und stellt danach
+auch bei einem Fehler das vorherige Profil wieder her. Das Ergebnis liegt in
+`/data/qwen-image-2.1-test-output/`. FLUX-Konfiguration und FLUX-Gewichte
+werden nicht verändert.
+
+## Vollständig entfernen
+
+```bash
+cd /opt/mike-ai/stack
+docker compose --env-file /etc/mike-ai/stack.env \
+  --profile qwen-image-test rm -sf qwen-image-21-test
+docker image rm mike-ai/qwen-image-2.1-test:local
+rm -rf /data/models/Qwen-Image-2.1-ComfyUI \
+  /data/qwen-image-2.1-test-output
+```
+
+Anschließend kann die Controller-Erweiterung bei Bedarf aus Git zurückgenommen
+und nur `profile-controller` neu gebaut werden. Die Entfernung ist nicht Teil
+der Vorbereitung und wird niemals automatisch ausgeführt.
@@ -27,6 +27,8 @@ LABEL_KEY = "com.mike-ai.llama-profile"
 IMAGE_LABEL_KEY = "com.mike-ai.image-worker"
 IMAGE_WORKER = os.environ.get("IMAGE_WORKER", "image")
 RESTORE_WORKER = os.environ.get("RESTORE_WORKER", "restore")
+QWEN_IMAGE_TEST_WORKER = os.environ.get(
+    "QWEN_IMAGE_TEST_WORKER", "qwen-image-2.1-test").strip()
 TTS_LABEL_KEY = "com.mike-ai.tts-worker"
 TTS_WORKER = os.environ.get("TTS_WORKER", "qwen3")
 MUSIC_LABEL_KEY = "com.mike-ai.music-worker"
@@ -101,6 +103,8 @@ def image_container(kind: str = IMAGE_WORKER) -> dict:
 def image_containers() -> list[dict]:
    """All allowlisted GPU workers that must never overlap an LLM."""
    allowed = {IMAGE_WORKER, RESTORE_WORKER}
+    if QWEN_IMAGE_TEST_WORKER:
+        allowed.add(QWEN_IMAGE_TEST_WORKER)
    return [item for item in labelled_containers(IMAGE_LABEL_KEY)
            if item.get("Labels", {}).get(IMAGE_LABEL_KEY) in allowed]

@@ -359,7 +363,7 @@ def stop_inference() -> dict:


 def set_image_worker(running: bool, kind: str = IMAGE_WORKER) -> dict:
-    if kind not in {IMAGE_WORKER, RESTORE_WORKER}:
+    if kind not in {IMAGE_WORKER, RESTORE_WORKER, QWEN_IMAGE_TEST_WORKER}:
        raise ValueError("worker is not allowlisted")
    with LOCK:
        item = image_container(kind)
@@ -797,6 +801,8 @@ class Handler(BaseHTTPRequestHandler):
            "/workers/image/stop": (IMAGE_WORKER, False),
            "/workers/restore/start": (RESTORE_WORKER, True),
            "/workers/restore/stop": (RESTORE_WORKER, False),
+            "/workers/qwen-image-test/start": (QWEN_IMAGE_TEST_WORKER, True),
+            "/workers/qwen-image-test/stop": (QWEN_IMAGE_TEST_WORKER, False),
        }
        if self.path in worker_paths:
            try:
@@ -0,0 +1,15 @@
+ARG PYTORCH_IMAGE=pytorch/pytorch:2.11.0-cuda12.8-cudnn9-runtime
+FROM ${PYTORCH_IMAGE}
+
+ARG COMFYUI_COMMIT
+RUN test -n "$COMFYUI_COMMIT" \
+    && apt-get update \
+    && apt-get install -y --no-install-recommends git ca-certificates \
+    && git clone --filter=blob:none https://github.com/Comfy-Org/ComfyUI.git /opt/ComfyUI \
+    && git -C /opt/ComfyUI checkout "$COMFYUI_COMMIT" \
+    && python -m pip install --no-cache-dir -r /opt/ComfyUI/requirements.txt \
+    && apt-get purge -y --auto-remove git \
+    && rm -rf /var/lib/apt/lists/* /root/.cache
+
+WORKDIR /opt/ComfyUI
+EXPOSE 8188
@@ -0,0 +1,35 @@
+#!/usr/bin/env bash
+# Download pinned official weights, build the isolated worker and leave it stopped.
+set -Eeuo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+ENV_FILE=${STACK_ENV:-/etc/mike-ai/stack.env}
+MODEL_DIR=${QWEN_IMAGE_21_MODEL_DIR:-/data/models/Qwen-Image-2.1-ComfyUI}
+OUTPUT_DIR=${QWEN_IMAGE_21_OUTPUT_DIR:-/data/qwen-image-2.1-test-output}
+BASE=https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/ace0edeb3791a594ddfa36ed5f41a178a394e921
+
+download() {
+  local relative=$1 expected=$2 target="$MODEL_DIR/$1"
+  install -d -m 0755 "$(dirname "$target")"
+  if [[ -f $target ]] && echo "$expected  $target" | sha256sum -c --status; then
+    echo "OK: $relative"
+    return
+  fi
+  curl -fL --retry 5 --retry-all-errors -C - -o "$target.part" "$BASE/$relative"
+  echo "$expected  $target.part" | sha256sum -c --status
+  mv "$target.part" "$target"
+}
+
+download diffusion_models/qwen_image_2.1_int8_convrot.safetensors cb74113cb03faecd79611b01fd7fd642f0aa60d6f0b95086abee214d75eaa57d
+download text_encoders/qwen3vl_8b_int8_convrot.safetensors 8bfd0f6e12abf2d2d697ecc888e5e90b0d6741d6708f05799f53afa560452e8f
+download vae/qwen_image_2.1_vae_bf16.safetensors bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9
+install -d -m 0755 "$OUTPUT_DIR"
+
+compose=(docker compose --env-file "$ENV_FILE" -f "$ROOT_DIR/compose.yaml" --profile qwen-image-test)
+"${compose[@]}" build qwen-image-21-test
+"${compose[@]}" up -d --build --no-deps profile-controller
+"${compose[@]}" create qwen-image-21-test
+"${compose[@]}" stop qwen-image-21-test
+state=$(docker inspect -f '{{.State.Status}}' mike-ai-qwen-image-2.1-test)
+[[ $state == exited || $state == created ]]
+echo "Qwen-Image-2.1 test is prepared and stopped. No GPU model was loaded."
@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+"""One-shot Qwen-Image-2.1 evaluation with guaranteed profile restoration."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import pathlib
+import random
+import time
+import urllib.parse
+import urllib.request
+
+
+CONTROLLER = os.environ.get("CONTROLLER_URL", "http://profile-controller:8090")
+TOKEN = os.environ["CONTROLLER_TOKEN"]
+COMFY = os.environ.get("COMFY_URL", "http://qwen-image-21-test:8188")
+OUTPUT = pathlib.Path(os.environ.get("OUTPUT_DIR", "/output"))
+
+
+def request_json(url: str, *, payload: dict | None = None,
+                 authenticated: bool = False, timeout: float = 30) -> dict:
+    body = None if payload is None else json.dumps(payload).encode()
+    headers = {"Content-Type": "application/json"}
+    if authenticated:
+        headers["Authorization"] = f"Bearer {TOKEN}"
+    method = "POST" if payload is not None else "GET"
+    req = urllib.request.Request(url, data=body, headers=headers, method=method)
+    with urllib.request.urlopen(req, timeout=timeout) as response:
+        return json.load(response)
+
+
+def controller(path: str, *, post: bool = False) -> dict:
+    return request_json(CONTROLLER + path, payload={} if post else None,
+                        authenticated=True, timeout=900)
+
+
+def wait_comfy(timeout: int = 900) -> None:
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        try:
+            request_json(COMFY + "/system_stats", timeout=3)
+            return
+        except Exception:
+            time.sleep(2)
+    raise TimeoutError("Qwen-Image worker did not become ready")
+
+
+def workflow(prompt: str, seed: int, steps: int, width: int, height: int) -> dict:
+    return {
+        "1": {"class_type": "UNETLoader", "inputs": {
+            "unet_name": "qwen_image_2.1_int8_convrot.safetensors",
+            "weight_dtype": "default"}},
+        "2": {"class_type": "CLIPLoader", "inputs": {
+            "clip_name": "qwen3vl_8b_int8_convrot.safetensors",
+            "type": "qwen_image", "device": "default"}},
+        "3": {"class_type": "VAELoader", "inputs": {
+            "vae_name": "qwen_image_2.1_vae_bf16.safetensors"}},
+        "4": {"class_type": "TextEncodeQwenImage21", "inputs": {
+            "clip": ["2", 0], "prompt": prompt, "negative_prompt": "",
+            "resolution": max(width, height)}},
+        "5": {"class_type": "EmptyLatentImage", "inputs": {
+            "width": width, "height": height, "batch_size": 1}},
+        "6": {"class_type": "KSampler", "inputs": {
+            "model": ["1", 0], "positive": ["4", 0], "negative": ["4", 1],
+            "latent_image": ["5", 0], "seed": seed, "steps": steps,
+            "cfg": 1.0, "sampler_name": "euler", "scheduler": "simple",
+            "denoise": 1.0}},
+        "7": {"class_type": "VAEDecode", "inputs": {
+            "samples": ["6", 0], "vae": ["3", 0]}},
+        "8": {"class_type": "SaveImage", "inputs": {
+            "filename_prefix": "qwen-image-2.1-test", "images": ["7", 0]}},
+    }
+
+
+def wait_result(prompt_id: str, timeout: int = 3600) -> dict:
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        history = request_json(f"{COMFY}/history/{prompt_id}", timeout=10)
+        if prompt_id in history:
+            result = history[prompt_id]
+            status = result.get("status", {})
+            if status.get("status_str") == "error" or not status.get("completed", True):
+                raise RuntimeError("generation failed: " + json.dumps(status))
+            return result
+        time.sleep(2)
+    raise TimeoutError("Qwen-Image generation timed out")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("prompt")
+    parser.add_argument("--seed", type=int, default=None)
+    parser.add_argument("--steps", type=int, default=25)
+    parser.add_argument("--width", type=int, default=1024)
+    parser.add_argument("--height", type=int, default=1024)
+    args = parser.parse_args()
+    if args.width % 32 or args.height % 32:
+        parser.error("width and height must be multiples of 32")
+    seed = args.seed if args.seed is not None else random.randrange(2**53)
+    previous = controller("/profiles/status").get("active_profile")
+    started = time.monotonic()
+    try:
+        controller("/workers/qwen-image-test/start", post=True)
+        wait_comfy()
+        queued = request_json(COMFY + "/prompt", payload={
+            "prompt": workflow(args.prompt, seed, args.steps, args.width, args.height),
+            "client_id": "athena-qwen-image-21-test"}, timeout=30)
+        prompt_id = queued["prompt_id"]
+        result = wait_result(prompt_id)
+        images = []
+        for node in result.get("outputs", {}).values():
+            images.extend(node.get("images", []))
+        if not images:
+            raise RuntimeError("generation completed without an image")
+        image = images[0]
+        query = urllib.parse.urlencode({
+            "filename": image["filename"], "subfolder": image.get("subfolder", ""),
+            "type": image.get("type", "output")})
+        target = OUTPUT / image["filename"]
+        target.parent.mkdir(parents=True, exist_ok=True)
+        with urllib.request.urlopen(COMFY + "/view?" + query, timeout=120) as src:
+            target.write_bytes(src.read())
+        print(json.dumps({"status": "ok", "file": str(target), "seed": seed,
+                          "seconds": round(time.monotonic() - started, 1)}, indent=2))
+    finally:
+        try:
+            controller("/workers/qwen-image-test/stop", post=True)
+        finally:
+            if previous:
+                controller(f"/profiles/{previous}/activate", post=True)
+
+
+if __name__ == "__main__":
+    main()