Any guidance / suggestion for B200 optimization?

[nanwan03]

Hi,

Thanks for the amazing work. While reviewing the documentation at LightX2V Readme
, we noticed there are no optimizations mentioned for B200 GPUs, particularly around model quantization and attention mechanisms.

Could you share any guidance or best practices for optimizing performance on B200 (sm100) GPUs? Specifically, we’d love to learn more about:

Recommended quantization settings or workflows
Attention optimizations (e.g. FlashAttention, SageAttention, or other kernels)
Known limitations or compatibility notes for sm100 architectures

Thanks.

[XHPlus]

Thanks for the question. We indeed has not tested on B200 due to the availability. Althought 5090 shares similar architecture, the best practice might differs. So it would be better to recommend after profiling on B200. If possible, we can make some tests on your machine and explores the best practice.

[nanwan03]

Thanks for the question. We indeed has not tested on B200 due to the availability. Althought 5090 shares similar architecture, the best practice might differs. So it would be better to recommend after profiling on B200. If possible, we can make some tests on your machine and explores the best practice.

Hi @XHPlus,

How many B200 GPUs would you need, and how long do you expect the testing to take?

My email address is nanw@reverie-ai.com. We can discuss the details further over email.

Best,
Nan Wang