cleanup-hf-model.yaml

x-logging-conf: &logging-conf
  driver: "json-file"
  options:
    max-size: "10m"
    max-file: "3"
    labels: "com.datadoghq.ad.logs"

# One-shot cleanup: removes one or more models from an existing HuggingFace
# cache volume on a CVM.
#
# Recommended usage (compose-manager `compose/up`):
#   curl -sS 'http://<host_ip>:8080/compose/up' \
#     -H 'Authorization: Bearer <CM_TOKEN>' -H 'Content-Type: application/json' \
#     -d '{"tag":"<tag>","file":"cleanup-hf-model.yaml","project":"cleanup",
#          "force_recreate":true,
#          "env":{"MODEL_NAME":"zai-org/GLM-5-FP8, Qwen/Qwen3.6-27B-FP8",
#                 "HF_CACHE_VOLUME":"work_hugginface_cache"}}'
#
# Notes:
# - MODEL_NAME accepts one or more `org/repo` values separated by commas and/or
#   whitespace and maps each to its on-disk `models--org--repo` directory.
# - HF_CACHE_VOLUME selects an EXISTING Docker volume by its full name (default:
#   `work_huggingface_cache`). The external declaration makes Compose fail fast
#   instead of silently creating an empty project-scoped volume.
# - Older `small-models.yaml` / `dsv4-qwen36-gemma4.yaml` packs produced the
#   misspelled `work_hugginface_cache` volume; select it explicitly when needed.
# - WARNING: Running without `"project":"cleanup"` uses the default `work`
#   project, where compose-manager's `--remove-orphans` will stop live model
#   containers. The separate `cleanup` project avoids touching the live stack.
# - Safe to run while inference is up: weights are already loaded into GPU
#   memory; only a future restart would need to re-download.
# - `force_recreate:true` is required to re-run on the same instance with a
#   different MODEL_NAME (otherwise the previous Exited container is reused).
services:
  hf-cleanup:
    image: alpine:3.20
    container_name: hf-cleanup
    restart: "no"
    entrypoint: ["sh", "-c"]
    command:
      - |
        set -eu
        : "$${MODEL_NAME:?MODEL_NAME env var is required (e.g. zai-org/GLM-5-FP8)}"

        echo "HF_CACHE_VOLUME=$$HF_CACHE_VOLUME"

        set -f
        MODELS=$$(printf '%s' "$$MODEL_NAME" | tr ',' ' ')
        set -- $$MODELS
        if [ "$$#" -eq 0 ]; then
          echo "Invalid MODEL_NAME: at least one org/repo is required." >&2
          exit 1
        fi
        for MODEL in "$$@"; do
          case "$$MODEL" in
            /*|*/|*/*/*|*..*)
              echo "Invalid model token '$$MODEL': expected exactly one org/repo with no leading slash or '..'." >&2
              exit 1
              ;;
            */*) ;;
            *)
              echo "Invalid model token '$$MODEL': expected exactly one org/repo with no leading slash or '..'." >&2
              exit 1
              ;;
          esac
        done

        echo "=== Disk usage ==="
        df -h /root/.cache 2>/dev/null || df -h /

        echo ""
        echo "=== Cached model weights ==="
        HUB=/root/.cache/huggingface/hub
        if ls "$$HUB"/models--* >/dev/null 2>&1; then
          du -sh "$$HUB"/models--* | sort -rh
        else
          echo "(none)"
        fi
        echo ""
        echo "Total HF cache: $$(du -sh $$HUB 2>/dev/null | cut -f1)"

        echo ""
        echo "=== Cleanup ==="
        TOTAL=$$#
        REMOVED=0
        for MODEL in "$$@"; do
          SAFE=$$(printf '%s' "$$MODEL" | sed 's|/|--|g')
          TARGET="$$HUB/models--$$SAFE"
          echo "MODEL_NAME=$$MODEL"
          echo "TARGET=$$TARGET"
          if [ ! -d "$$TARGET" ]; then
            echo "Not found — nothing to remove."
            continue
          fi
          echo "Removing $$(du -sh "$$TARGET" | cut -f1) ..."
          rm -rf "$$TARGET"
          REMOVED=$$((REMOVED + 1))
          echo "Removed $$TARGET"
        done
        echo "Disk after: $$(df -h /root/.cache 2>/dev/null | tail -1 || df -h / | tail -1)"
        echo "Removed $$REMOVED of $$TOTAL requested model(s)."
    volumes:
      - huggingface_cache:/root/.cache/huggingface
    environment:
      - MODEL_NAME=${MODEL_NAME}
      - HF_CACHE_VOLUME=${HF_CACHE_VOLUME:-work_huggingface_cache}
    logging: *logging-conf

networks:
  default:
    external: true
    name: dstack_default

volumes:
  huggingface_cache:
    external: true
    name: ${HF_CACHE_VOLUME:-work_huggingface_cache}