# Serve scientific_granite with vLLM (OpenAI-compatible /v1/embeddings).
#
#   docker compose -f compose.vllm.yml up -d      # needs an NVIDIA GPU + NVIDIA Container Toolkit
#
# Then:
#   curl localhost:8000/v1/embeddings -H 'Content-Type: application/json' \
#     -d '{"model": "daniel-gomm/scientific_granite", "input": ["your query"]}'
#
# vLLM returns unnormalized embeddings for this model: L2-normalize before dot-product
# similarity (cosine is unaffected). GPU memory: ~0.5 GB; the model needs no KV cache, so a
# small --gpu-memory-utilization leaves the rest of the GPU free.
name: scientific-granite-vllm

services:
  vllm:
    build:
      # orjson makes vLLM's /v1/embeddings JSON encoding much faster (vLLM warns without it)
      dockerfile_inline: |
        FROM vllm/vllm-openai:${VLLM_VERSION:-v0.31.0}
        RUN pip install --no-cache-dir orjson
    image: scientific-granite-vllm:${VLLM_VERSION:-v0.31.0}
    entrypoint: ["vllm", "serve"]
    command: >-
      ${MODEL_ID:-daniel-gomm/scientific_granite}
      --runner pooling
      --pooler-config '{"pooling_type": "CLS"}'
      --max-model-len 2048
      --gpu-memory-utilization ${GPU_MEM_UTIL:-0.25}
      --disable-uvicorn-access-log
    ports:
      - "${PORT:-8000}:8000"
    ipc: host
    volumes:
      - hf-cache:/root/.cache/huggingface  # downloaded model files
      - vllm-cache:/root/.cache/vllm       # torch.compile artifacts: faster restarts
    environment:
      HF_TOKEN: ${HF_TOKEN:-}
    healthcheck:
      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
      interval: 10s
      start_period: 120s
    restart: unless-stopped
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

volumes:
  hf-cache:
  vllm-cache: