| services: | |
| server: | |
| build: | |
| context: . | |
| # Enable BuildKit for better caching | |
| cache_from: | |
| - python:3.9 | |
| ports: | |
| # Bind to loopback only by default: this service has no authentication (see README → | |
| # "Deployment & network"). Change to "7860:7860" (all interfaces) or attach it to an | |
| # internal Docker network only after you have placed it behind your own trust boundary. | |
| - 127.0.0.1:7860:7860 | |
| environment: | |
| # Local inference service: models are pre-cached in the python-cache volume, so load from | |
| # cache and never revalidate against huggingface.co at runtime (avoids cold-start hangs when | |
| # the network to HF is slow/unavailable). Remove/set to 0 only when adding a new model. | |
| HF_HUB_OFFLINE: "1" | |
| TRANSFORMERS_OFFLINE: "1" | |
| # Namespaces whose models may run their own Python at load time (trust_remote_code). | |
| # Everything else still loads — just with remote code disabled, which covers every | |
| # architecture transformers supports natively. Add an org only when its models genuinely | |
| # need custom code; each entry is a standing grant to everyone that org ever adds. | |
| # On the Hugging Face Space this is set as a Space Variable instead of here. | |
| TRUSTED_CODE_ORGS: "nomic-ai,pimcore" | |
| # How many loaded models stay resident (least-recently-used evicted first). Counts models, | |
| # not bytes — lower it when serving large vision models. Also a Space Variable on HF. | |
| MAX_CACHED_MODELS: "3" | |
| develop: | |
| watch: | |
| # Only rebuild on requirements.txt changes, sync code changes otherwise | |
| - action: rebuild | |
| path: ./requirements.txt | |
| - action: sync | |
| path: ./src | |
| target: /home/user/app/src | |
| - action: sync | |
| path: ./README.md | |
| target: /home/user/app/README.md | |
| volumes: | |
| - python-cache:/home/user/.cache | |
| # Cache pip packages | |
| - pip-cache:/home/user/.cache/pip | |
| volumes: | |
| python-cache: | |
| pip-cache: |