Models¶

Jeff serves GLiFormer checkpoints. The image does not contain weights: they are downloaded on first start into the /models volume.

Model	Params	Disk	Notes
`knowledgator/gliformer-large-v1`	~400 M	~1.7 GB	Default, best quality
`knowledgator/gliformer-base-v1`	~264 M	~1.1 GB	Faster, lower quality; set `JEFF_NOUL_MODE=single`

How the download works¶

On start the entrypoint looks for <JEFF_MODEL_PATH>/.jeff-download-complete.
Missing → hf download <JEFF_MODEL_ID> --revision <JEFF_MODEL_REVISION> into JEFF_MODEL_PATH, then the marker is written.
Present → download skipped (Model already present in the logs).

An interrupted download leaves no marker, so it resumes on the next start.

Switching to the base model¶

JEFF_MODEL_ID=knowledgator/gliformer-base-v1
JEFF_MODEL_PATH=/models/gliformer-base-v1
JEFF_NOUL_MODE=single
JEFF_MODEL_NAME=gliformer-base-v1

Both models can coexist in the same volume (different paths).

Pinning a revision (reproducible deployments)¶

curl -s https://huggingface.co/api/models/knowledgator/gliformer-large-v1 | jq -r .sha

JEFF_MODEL_REVISION=<that sha>
JEFF_FORCE_MODEL_DOWNLOAD=true   # once, to replace the current weights

Set JEFF_FORCE_MODEL_DOWNLOAD back to false after the first start.

Private or gated repositories¶

HF_TOKEN=hf_xxx
JEFF_MODEL_ID=your-org/your-finetune
JEFF_MODEL_PATH=/models/your-finetune

Offline / air-gapped hosts¶

Download on a connected machine, copy the folder, mount it read-only:

# connected machine
pip install -U "huggingface_hub[cli]"
hf download knowledgator/gliformer-large-v1 --local-dir ./gliformer-large-v1
tar czf gliformer-large-v1.tgz gliformer-large-v1

# offline host
sudo mkdir -p /srv/jeff/models && sudo tar xzf gliformer-large-v1.tgz -C /srv/jeff/models
sudo chown -R 10001:10001 /srv/jeff/models

volumes:
  - /srv/jeff/models:/models
environment:
  JEFF_DOWNLOAD_MODEL: "false"
  HF_HUB_OFFLINE: "1"

With JEFF_DOWNLOAD_MODEL=false, a non-empty directory without marker is accepted as manually provisioned.

Pre-seeding a named volume¶

docker run --rm -v jeff-models:/models -e JEFF_API_KEYS=x \
  --entrypoint hf ghcr.io/tommasomarchionni/jeff-docker:latest \
  download knowledgator/gliformer-large-v1 --local-dir /models/gliformer-large-v1
docker run --rm -v jeff-models:/models --entrypoint sh \
  ghcr.io/tommasomarchionni/jeff-docker:latest \
  -c 'echo manual > /models/gliformer-large-v1/.jeff-download-complete'

Licenses¶

Check each model card on Hugging Face for its license before commercial use. This repository's MIT license covers the packaging only.