backup stack: whisper-cpp-sycl [2026-11-05]

stacks/4_whisper_cpp_sycl.yml

@@ -0,0 +1,57 @@
+services:
+  # One-shot helper: downloads models into the models folder if missing, then exits.
+  whisper-models:
+    image: alpine:3
+    container_name: whisper-models
+    restart: "no"
+    volumes:
+      - /home/fireball/docker/whisper-cpp/models:/models
+    command:
+      - sh
+      - -c
+      - |
+        set -e
+        apk add --no-cache wget >/dev/null
+        get() {
+          if [ -s "/models/$$1" ]; then echo "have $$1"; else
+            echo "downloading $$1"
+            wget -q --show-progress -O "/models/$$1.part" "$$2" && mv "/models/$$1.part" "/models/$$1"
+          fi
+        }
+        get ggml-large-v3.bin        https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3.bin
+        get ggml-silero-v5.1.2.bin   https://huggingface.co/ggml-org/whisper-vad/resolve/main/ggml-silero-v5.1.2.bin
+
+  whisper-sycl:
+    image: ghcr.io/linktothedata/whisper-sycl:latest
+    container_name: whisper-sycl
+    restart: unless-stopped
+    depends_on:
+      whisper-models:
+        condition: service_completed_successfully
+    ports:
+      - "8081:8080"            # whisper-server UI + /inference
+    devices:
+      - /dev/dxg:/dev/dxg
+      # - /dev/dri:/dev/dri    # add if your ollama-sycl service maps it
+    volumes:
+      - /usr/lib/wsl:/usr/lib/wsl   # WSL GPU libs (match ollama-sycl)
+      - /home/fireball/docker/whisper-cpp/models:/models
+    environment:
+      - ONEAPI_DEVICE_SELECTOR=level_zero:0   # force the Arc, not the CPU
+      - SYCL_CACHE_PERSISTENT=1               # faster restarts
+    # Image entrypoint may be the CLI, so find whisper-server explicitly.
+    entrypoint: ["/bin/sh", "-c"]
+    command:
+      - >
+        exec "$$(find / -name whisper-server -type f 2>/dev/null | head -1)"
+        --host 0.0.0.0 --port 8080
+        -m /models/ggml-large-v3.bin
+        --convert --language auto
+        --vad -vm /models/ggml-silero-v5.1.2.bin
+    healthcheck:
+      # bash /dev/tcp avoids needing curl in the image
+      test: ["CMD-SHELL", "bash -c '</dev/tcp/127.0.0.1/8080' || exit 1"]
+      interval: 30s
+      timeout: 5s
+      retries: 3
+      start_period: 120s