version: "3.9"

# ── Local development / testing compose file ─────────────────────────────────
# For HuggingFace Spaces deployment, only the Dockerfile is needed.
#
# Usage:
#   docker compose build
#   HF_TOKEN=hf_xxx docker compose up

services:
  converter:
    build: .
    image: dsmath-gguf-converter:latest
    container_name: dsmath-convert

    # Mount a local directory as the workspace so artefacts persist
    volumes:
      - ./workspace:/workspace

    environment:
      # Set your HF token here or export it in your shell
      - HF_TOKEN=${HF_TOKEN:-}

      # Skip flags – useful when resuming a partial run
      - SKIP_DOWNLOAD=${SKIP_DOWNLOAD:-0}
      - SKIP_FP8_CAST=${SKIP_FP8_CAST:-0}
      - SKIP_GGUF_CONVERT=${SKIP_GGUF_CONVERT:-0}

      # Override paths if needed
      - FP8_DIR=/workspace/model-fp8
      - BF16_DIR=/workspace/model-bf16
      - GGUF_BF16=/workspace/model-gguf/model-bf16.gguf
      - GGUF_Q8=/workspace/output/deepseek-math-v2-q8_0.gguf
      - LLAMA_CPP_DIR=/opt/llama.cpp

    # The main conversion pipeline
    command: bash /workspace/scripts/convert.sh

    # Resource hints (adjust to your machine)
    deploy:
      resources:
        limits:
          memory: 800g