Sharing and export

Save local folders with save_pretrained, publish with push_to_hub, and export when a production runtime needs ONNX, ExecuTorch or another format.

Adapted for this playbook from the 🤗 Transformers documentation by Hugging Face. Official page: https://huggingface.co/docs/transformers/model_sharing. Images © Hugging Face (documentation-images) unless noted. This is not a substitute for the upstream docs — verify against the current version.

Covers Hugging Face pages: model_sharing, serialization

Source: Hugging Face documentation images.

Save locally

model.save_pretrained("./my-model")
tokenizer.save_pretrained("./my-model")

Push to the Hub

model.push_to_hub("org/my-model")
tokenizer.push_to_hub("org/my-model")

Authenticate with huggingface-cli login or HF_TOKEN. Add a model card (README.md) covering intended use, limitations and evaluation.

Source: Hugging Face documentation images.

Exporting for production

Transformers documents exporters and serialisation paths for moving beyond eager PyTorch — see Exporting to production and the Exporters guide. Typical targets:

• ONNX / Optimum for portable graphs
• Runtime-specific formats (TensorRT, ExecuTorch, llama.cpp GGUF via conversion tools)
• Serve with engines that load HF definitions directly (vLLM)

Enterprise checklist

• Licence compatible with redistribution
• Model card and eval evidence attached
• Secrets and training data not accidentally uploaded
• Revision tag created for the release

Key takeaways

• save_pretrained / push_to_hub are the standard share path.
• Export only when the serving stack requires it — many engines consume HF folders natively.

What to do next

Tokenizers or Inference engines.