Deployment recommendation

[zyandtom]

Do you have any advice on deployment? It seems that frameworks like vLLM and SGLang aren't supported at the moment.

[Robin970822]

Thanks for raising this question.

At the moment, our officially supported deployment path is the PyTorch reference implementation in the repository:

• Inference guide
• Inference wrapper
• Gradio demo

We have not yet validated SenseNova-Vision with vLLM or SGLang, so we cannot currently recommend either as a production serving backend.

There is an important distinction regarding BAGEL support:

• Standard vLLM BAGEL support focuses on vision-to-text inference and does not support BAGEL's image-generation path.
• SenseNova-Vision uses image outputs for tasks such as segmentation, depth, surface normals, and multi-view reconstruction, so standard vLLM support would not cover the full model capability.
• vLLM-Omni has broader support for BAGEL-related pipelines and may be a relevant direction, but we have not yet tested the SenseNova-Vision checkpoint and all of its task modes with it.

For now, please use the repository's reference inference path. It is intended for functional inference and evaluation, rather than high-concurrency production serving.

We are tracking deployment support, but do not have a committed timeline yet. If you can share your target tasks, GPU configuration, expected concurrency, and latency requirements, that would help us understand and prioritize the use case.