About GPU requirement

[Lyman-Smoker]

Thank you for your excellent work on this project!

I’m interested in training the model on my side, and I am curious about minimum GPU memory required for training.

Can the model be trained on GPUs with less memory (e.g, 8 * 40G)?

Thanks for your help in advance!

[oneScotch]

Hi, thank you for your interest in our work!

For reference, we trained the model using 8 × H100 GPUs ( 8 x 80 GB).

Training on GPUs with less memory may still be possible, but it would likely require adjustments such as reducing the batch size, enabling gradient checkpointing, or using more aggressive memory optimization techniques. The exact feasibility depends on the configuration you choose and the model size.

We haven’t extensively validated training under lower-memory setups yet, so some experimentation may be needed to find a stable configuration.

Hope this helps, and please feel free to reach out if you have further questions!