name: llm-embeddings

services:
  qwen3-embedding:
    image: ${EMBEDDING_IMAGE:-ghcr.io/ggml-org/llama.cpp@sha256:3e8914c1aab600a330ada97c10fb5fb02ff1c15ac39f2ece218db125cf54594e}
    container_name: ${EMBEDDING_CONTAINER_NAME:-llm-qwen3-embedding}
    restart: unless-stopped
    ports:
      - "${EMBEDDING_HOST_PORT:-8082}:8080"
    volumes:
      - ${HOST_LLAMA_CACHE_DIR:-Z:/LLM/models/cache/llama.cpp}:/root/.cache/llama.cpp
    command:
      - --host
      - 0.0.0.0
      - --port
      - "8080"
      - --hf-repo
      - ${EMBEDDING_HF_REPO:-Qwen/Qwen3-Embedding-0.6B-GGUF:Q8_0}
      - --alias
      - ${EMBEDDING_SERVED_MODEL_NAME:-qwen3-embedding-0.6b}
      - --embedding
      - --pooling
      - last
      - --ctx-size
      - ${EMBEDDING_CTX_SIZE:-8192}
      - --batch-size
      - ${EMBEDDING_BATCH_SIZE:-1024}
      - --ubatch-size
      - ${EMBEDDING_UBATCH_SIZE:-1024}
      - --threads
      - ${EMBEDDING_THREADS:-12}
      - --parallel
      - ${EMBEDDING_PARALLEL:-1}
      - --n-gpu-layers
      - "0"
    healthcheck:
      test:
        - CMD
        - curl
        - --fail
        - --silent
        - http://localhost:8080/health
      interval: 10s
      timeout: 5s
      retries: 30
      start_period: 10m
    deploy:
      resources:
        limits:
          cpus: ${EMBEDDING_CPU_LIMIT:-8.0}
          memory: ${EMBEDDING_MEMORY_LIMIT:-4G}