manage FireRed OCR with Arcane

servers/unraid/firered-ocr/README.md

@@ -0,0 +1,11 @@
+# FireRed OCR
+
+Arcane-managed FireRed-OCR vLLM endpoint on unRaid GPU1.
+
+- API: `http://192.168.50.100:8001`
+- Model data: `/mnt/cache-appdata/firered-ocr`
+- GPU: physical device 1, exposed as CUDA device 0
+- Image: pinned to the digest already present on unRaid
+
+The service was migrated from the `llama_cpp` Compose Manager project. The
+model bind mount is reused in place; no model data is copied into Arcane.

servers/unraid/firered-ocr/compose.yaml

@@ -0,0 +1,47 @@
+services:
+  firered-ocr:
+    image: vllm/vllm-openai@sha256:2296a2a7e1ce1dc59c6577ba5900f4e9910b76c4a0cb134833a8137f92404dfa
+    pull_policy: never
+    container_name: firered-ocr
+    restart: unless-stopped
+    entrypoint: ["/bin/sh", "-c"]
+    command:
+      - >-
+        pip install qwen-vl-utils -q &&
+        vllm serve /models/firered-ocr
+        --served-model-name firered-ocr
+        --load-format safetensors
+        --trust-remote-code
+        --dtype half
+        --gpu-memory-utilization 0.30
+        --max-num-seqs 1
+        --max-model-len 4096
+        --enforce-eager
+        --host 0.0.0.0
+        --port 8001
+    environment:
+      NVIDIA_VISIBLE_DEVICES: "1"
+      CUDA_VISIBLE_DEVICES: "0"
+      NVIDIA_DRIVER_CAPABILITIES: compute,utility
+    ports:
+      - "8001:8001"
+    volumes:
+      - /mnt/cache-appdata/firered-ocr:/models/firered-ocr
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ["1"]
+              capabilities: [gpu]
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8001/health"]
+      interval: 30s
+      timeout: 10s
+      start_period: 90s
+      retries: 3
+    labels:
+      com.getarcaneapp.arcane.icon: firered-ocr
+      net.unraid.docker.icon: https://github.com/FireRedTeam.png?size=128
+      net.unraid.docker.webui: http://[IP]:[PORT:8001]/docs
+