Simplify Athena runtime and document current architecture

platform/docker/image-worker/Dockerfile

@@ -7,15 +7,15 @@ ARG HF_HUB_VERSION=1.28.0
 
 RUN apt-get update && apt-get install -y --no-install-recommends python3.12-venv && \
     rm -rf /var/lib/apt/lists/* && \
-    python -m venv --system-site-packages /opt/flux-venv && \
-    /opt/flux-venv/bin/pip install --no-cache-dir \
+    python -m venv --system-site-packages /opt/image-venv && \
+    /opt/image-venv/bin/pip install --no-cache-dir \
       "diffusers==${DIFFUSERS_VERSION}" \
       "transformers==${TRANSFORMERS_VERSION}" \
       "accelerate==${ACCELERATE_VERSION}" \
       "huggingface-hub==${HF_HUB_VERSION}" \
       sentencepiece protobuf safetensors pillow && \
-    useradd --system --uid 10002 --home /nonexistent --shell /usr/sbin/nologin flux
+    useradd --system --uid 10002 --home /nonexistent --shell /usr/sbin/nologin image-worker
 
-COPY flux_worker.py /app/flux_worker.py
+COPY image_worker.py /app/image_worker.py
 USER 10002:10002
-ENTRYPOINT ["/opt/flux-venv/bin/python", "/app/flux_worker.py"]
+ENTRYPOINT ["/opt/image-venv/bin/python", "/app/image_worker.py"]

platform/docker/image-worker/image_worker.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Private FLUX.2 Klein Distilled worker used only during a GPU hot swap."""
+"""Private Z-Image-Turbo worker used only during a GPU hot swap."""
 
 from __future__ import annotations
 
@@ -14,7 +14,7 @@ from pathlib import Path
 HOST = os.environ.get("WORKER_HOST", "0.0.0.0")
 PORT = int(os.environ.get("WORKER_PORT", "8086"))
 TOKEN = os.environ.get("WORKER_TOKEN", "").strip()
-MODEL_DIR = os.environ.get("FLUX_MODEL_DIR", "/models/FLUX.2-klein-4B")
+MODEL_DIR = os.environ.get("Z_IMAGE_MODEL_DIR", "/models/Z-Image-Turbo")
 OUTPUT_DIR = Path(os.environ.get("IMAGE_DIR", "/data/images")).resolve()
 PIPE = None
 LOAD_SECONDS = 0.0
@@ -33,14 +33,15 @@ def load_pipeline() -> None:
     if PIPE is not None:
         return
     import torch
-    from diffusers import DiffusionPipeline
+    from diffusers import ZImagePipeline
     started = time.monotonic()
-    PIPE = DiffusionPipeline.from_pretrained(
-        MODEL_DIR, torch_dtype=torch.bfloat16)
-    # The full pipeline leaves too little activation headroom on a 16 GiB
-    # RTX 5080. Model CPU offload keeps each active component on CUDA while
-    # parking inactive components in system RAM between the four steps.
-    PIPE.enable_model_cpu_offload()
+    PIPE = ZImagePipeline.from_pretrained(
+        MODEL_DIR, torch_dtype=torch.bfloat16, low_cpu_mem_usage=False)
+    # The Qwen text encoder and the DiT do not fit together in the usable
+    # 16 GiB of the RTX 5080. Sequential offload keeps only the active
+    # submodule on CUDA. This is slower than a fully resident pipeline, but
+    # deterministic and leaves the RTX 3060 available for XTTS.
+    PIPE.enable_sequential_cpu_offload()
     if hasattr(PIPE, "enable_vae_slicing"):
         PIPE.enable_vae_slicing()
     if hasattr(PIPE, "enable_vae_tiling"):
@@ -61,16 +62,16 @@ def generate(data: dict) -> dict:
     if (width, height) not in {(1024, 1024), (1536, 1024), (1024, 1536),
                                (1920, 1088), (1088, 1920)}:
         raise ValueError("unsupported image size")
-    steps = int(data.get("steps", 4))
-    guidance = float(data.get("guidance", 1.0))
-    if steps != 4 or guidance != 1.0:
-        raise ValueError("distilled FLUX.2 Klein requires steps=4 and guidance=1.0")
+    steps = int(data.get("steps", 9))
+    guidance = float(data.get("guidance", 0.0))
+    if steps != 9 or guidance != 0.0:
+        raise ValueError("Z-Image-Turbo requires steps=9 and guidance=0.0")
     seed = data.get("seed")
     generator = None if seed is None else torch.Generator(device="cuda").manual_seed(int(seed))
     load_pipeline()
     started = time.monotonic()
     image = PIPE(prompt=prompt, height=height, width=width,
-                 num_inference_steps=4, guidance_scale=1.0,
+                 num_inference_steps=9, guidance_scale=0.0,
                  generator=generator).images[0]
     OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
     output = OUTPUT_DIR / filename
@@ -83,7 +84,7 @@ def generate(data: dict) -> dict:
 class Handler(BaseHTTPRequestHandler):
     def log_message(self, fmt: str, *args: object) -> None:
         # Never log request bodies/prompts.
-        print(f"[flux-worker] {self.client_address[0]} {fmt % args}", flush=True)
+        print(f"[z-image-worker] {self.client_address[0]} {fmt % args}", flush=True)
 
     def reply(self, status: int, payload: dict) -> None:
         body = json.dumps(payload, separators=(",", ":")).encode()
@@ -112,7 +113,7 @@ class Handler(BaseHTTPRequestHandler):
                 raise ValueError("invalid request size")
             self.reply(200, generate(json.loads(self.rfile.read(length))))
         except Exception as exc:
-            print(f"[flux-worker] generation failed: "
+            print(f"[z-image-worker] generation failed: "
                   f"{type(exc).__name__}: {str(exc)[:1000]}", flush=True)
             self.reply(400, {"status": "error", "message": str(exc)})
 

platform/docker/profile-controller/profile_controller.py

@@ -24,7 +24,7 @@ ALLOWED = tuple(x.strip() for x in os.environ.get(
     "ALLOWED_PROFILES", "fast,medium,large,ultra,uncensored,experimental").split(",") if x.strip())
 LABEL_KEY = "com.mike-ai.llama-profile"
 IMAGE_LABEL_KEY = "com.mike-ai.image-worker"
-IMAGE_WORKER = os.environ.get("IMAGE_WORKER", "flux")
+IMAGE_WORKER = os.environ.get("IMAGE_WORKER", "image")
 LOCK = threading.Lock()
 log = logging.getLogger("profile-controller")
 
@@ -98,7 +98,7 @@ def set_image_worker(running: bool) -> dict:
     with LOCK:
         item = image_container()
         if running:
-            # A FLUX worker may never overlap a llama profile on the 5080.
+            # The image worker may never overlap a llama profile on the 5080.
             for profile_item in containers().values():
                 stop_container(profile_item)
             if item.get("State") != "running":

platform/docker/wireguard-gateway/entrypoint.sh

@@ -64,7 +64,7 @@ wg_ipv4=$(ip -4 -o address show dev wg0 | awk 'NR == 1 { split($4, address, "/")
 # The VPN is Athena's normal application network. Nothing below is published
 # on the physical university interface: every listener is bound inside this
 # namespace to the Fritzbox-assigned WireGuard address. Clients on the home
-# VPN may use OpenWebUI, the router and every useful MCP directly.
+# VPN clients may use the router, speech services and Athena operator directly.
 proxy_pids=""
 start_proxy() {
   listen_port=$1
@@ -74,29 +74,10 @@ start_proxy() {
 }
 
 start_proxy 22   172.30.10.1:22
-start_proxy 8080 open-webui:8080
 start_proxy 8081 router:8081
 start_proxy 8085 tts-gateway:8085
 start_proxy 8091 piper:8085
 start_proxy 8092 xtts:80
-start_proxy 9119 hermes:9119
-start_proxy 8642 hermes:8642
-
-# MCP endpoints. Optional services keep their listener even while stopped and
-# begin working automatically as soon as their container is started.
 start_proxy 8202 mcp-athena-operator:8000
-# Portable general web MCP for Pi, Hermes and other clients. OpenWebUI uses
-# its native broad search by default; both paths are site-agnostic.
-start_proxy 8203 tinysearch:8000
-start_proxy 8204 mcp-github:8000
-start_proxy 8205 mcp-homeassistant:8000
-start_proxy 8206 mcp-arr:8000
-start_proxy 8207 mcp-navidrome:3000
-start_proxy 8208 mcp-unraid-ssh:8000
-
-# Search backends are also directly available for diagnostics and alternative
-# clients. Normal chat clients should prefer the MCP endpoint on 8203.
-start_proxy 8210 searxng:8080
-start_proxy 8211 tinysearch:8000
 
 wait $(printf '%s\n' "$proxy_pids" | awk '{print $2}')