Add OneComp (Fujitsu) to Inference/Quantization

[Yuma-Ichikawa]

Summary

This PR adds OneComp (Fujitsu Research, arXiv:2603.28845) and its accompanying NeurIPS 2025 paper QEP to this awesome list.

OneComp is an open-source, Apache-2.0 licensed post-training quantization (PTQ) framework for LLMs that unifies several recent research directions behind a single Runner.auto_run() API:

• QEP (Quantization Error Propagation, NeurIPS 2025) — layer-wise PTQ with propagated error compensation.
• AutoBit — ILP-based mixed-precision bitwidth assignment derived from available VRAM.
• JointQ — joint weight–scale optimization (group-wise 4-bit, etc.).
• Rotation preprocessing — SpinQuant/OstQuant-style learned rotations absorbed into weights, with online Hadamard hooks at load time.
• LoRA SFT post-process — accuracy recovery / knowledge injection after quantization.
• vLLM plugin — first-class DBF & Mixed-GPTQ serving of OneComp checkpoints.

Verified on Llama (TinyLlama / Llama-2 / Llama-3) and Qwen3 (0.6B — 32B). PyPI: pip install onecomp.

Links:

• Repo: https://github.com/FujitsuResearch/OneCompression
• Paper (OneComp): https://arxiv.org/abs/2603.28845
• Paper (QEP, NeurIPS 2025): https://openreview.net/forum?id=a3l3K9khbL
• Docs: https://FujitsuResearch.github.io/OneCompression/

Happy to adjust section placement, wording, or formatting to better match the list's conventions — thanks for maintaining this resource!

[cr7258]

@Yuma-Ichikawa
Thanks for your contribution. Please use the Project Request issue template to submit a new project.
Here is an example: #351
The CI will update all necessary files.