Training time for world-r1

[HL-hanlin]

Thanks for the great work!

I'm currently trying to reproduce the results with Wan2.1-T2V-14B. Using the default single-node training script, I noticed that each epoch, including both sampling rollouts and training optimization steps, takes around 8–10 hours to finish. May I ask whether this is expected?

In addition, when scaling to multi-node training, it seems that the current training code also scales up the global batch size, while keeping num_batches_per_epoch the same for each rank. As a result, the per-epoch training time does not really decrease with more nodes. May I ask whether this behavior is intended?

Finally, if possible, could you share how many epochs the 1.3B and 14B models were trained for?

Thanks a lot!

[lhmd]

Hi @HL-hanlin .Thanks for your interest!

The default single-node training script is mainly provided as a reference implementation. Based on our experience, single-node training is usually not sufficient to obtain good/reproducible performance for this setting.

For reproducing the reported results, we recommend using the resource configuration described in the paper and training on the full dataset. Training for 1 epoch should be enough. Also, please keep the overall batch size unchanged, especially the on/off-policy update size, i.e. group_size * num_group, following the Flow-GRPO setting.

In practice, you should be able to observe meaningful results after training for a few dozen steps.