Toolkit overview

Eole is a PyTorch toolkit for training, fine-tuning, evaluating, and serving encoder, decoder, and encoder-decoder models. It originated from OpenNMT-py.

Start with the README and installation instructions, the quickstart, the Docker guide, or the recipe index. The reference sidebar is generated from the current Python configuration and API.

Current workflows include supported Hugging Face model conversion and direct inference, LoRA and quantized fine-tuning, translation scoring, vision/OCR, Whisper transcription, and an HTTP server with OpenAI-style and Anthropic-style endpoints. Hardware, precision, and optional kernels determine which paths run.

Qwen3.8 native MTP inference currently supports single-sequence greedy text requests. The recorded INT4/BF16 outputs differ from ordinary greedy decoding; exact production token parity is unresolved. Its vision-language training class loads MTP heads but does not implement their auxiliary training loss. REINFORCE with scorer rewards is implemented; DPO, GRPO, and PPO are planned. See the recipes for these limitations and what was actually tested.

Citation

For work building on Eole's OpenNMT lineage, cite the OpenNMT technical report:

@inproceedings{opennmt,
  author = {Guillaume Klein and Yoon Kim and Yuntian Deng and
            Jean Senellart and Alexander M. Rush},
  title = {OpenNMT: Open-Source Toolkit for Neural Machine Translation},
  booktitle = {Proceedings of ACL},
  year = {2017},
  doi = {10.18653/v1/P17-4012}
}

Use Discussions for questions and Issues for bugs.