github-action.yml

# Example: prewarm a Hugging Face cache from a Backblaze B2 bucket in CI.
#
# Drop this into .github/workflows/ in a repo that needs HF models at
# training/inference time. The shared cache cuts cold-start download
# time from ~15 minutes (HF hub) to ~30 seconds (regional B2 → runner).
#
# Required GitHub secrets:
#   HF_CACHE_BUCKET            — name of the B2 bucket
#   HF_CACHE_REGION            — bucket region (e.g. us-west-000)
#   HF_CACHE_ENDPOINT          — S3 endpoint URL (https://s3.<region>.backblazeb2.com)
#   HF_CACHE_ACCESS_KEY        — read-only B2 application key ID
#   HF_CACHE_SECRET_KEY        — read-only B2 application key

name: prewarm-hf-cache

on:
  workflow_call:
    inputs:
      models:
        description: "Space-separated list of repo IDs to pull (e.g. mistralai/Mistral-7B-v0.1)"
        type: string
        required: true

jobs:
  prewarm:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"

      - name: Install hf-cache-sync (with fallback extra)
        run: pip install "hf-cache-sync[fallback]"

      # The cache survives across runs on the same self-hosted runner; on
      # GitHub-hosted runners use actions/cache to persist ~/.cache/huggingface
      # between jobs.
      - name: Restore HF cache
        uses: actions/cache@v4
        with:
          path: ~/.cache/huggingface
          key: hf-${{ hashFiles('**/requirements*.txt') }}

      - name: Write hf-cache-sync config
        run: |
          mkdir -p ~/.config
          cat > ~/.hf-cache-sync.yaml <<EOF
          storage:
            endpoint: ${{ secrets.HF_CACHE_ENDPOINT }}
            bucket: ${{ secrets.HF_CACHE_BUCKET }}
            region: ${{ secrets.HF_CACHE_REGION }}
            access_key: ${{ secrets.HF_CACHE_ACCESS_KEY }}
            secret_key: ${{ secrets.HF_CACHE_SECRET_KEY }}
          EOF

      # Sanity check — fail fast if the bucket / creds are misconfigured.
      - name: Doctor
        run: hf-cache-sync doctor

      # Pull each model. --fallback hf-hub means if the team hasn't pushed
      # this model yet, CI still works (it pulls from HF hub directly).
      # Auth errors against B2 still surface — they don't get masked.
      - name: Pull models
        run: |
          for repo in ${{ inputs.models }}; do
            hf-cache-sync pull "$repo" --fallback hf-hub --workers 16
          done