Confusion Regarding max_step

[wwwjjjj]

I am confused about the max_step in this work: In the GRPO training documentation at https://github.com/modelscope/ms-swift/blob/3a33b7df5b8bb26982ee4f6c65b5fd4fb6b1813c/docs/source/BestPractices/GRPO%E5%A4%9A%E6%A8%A1%E6%80%81%E8%AE%AD%E7%BB%83.md, the author mentions training on 8k data for 1200+ steps. However, I observed that only 1 epoch was trained, and the batch size was 6 * 8 * 8. Shouldn't the max_step be calculated as 8000 / (6 * 8 * 8)?

[hjh0119]

Note: per_device_batch_size is completion-level, so when calculating the prompt-level batch_size, it needs to be divided by num_generations.

So based on the parameters of the script you mentioned:

total_prompt_data_size = 8000 * 8 (num_generations) / 8 (per_device_batch_size) / 6 (dp_size) * 0.99 (train_data_ratio) = 1320

max_step =total_prompt_data_size / 2 (ga_steps) * 2 (num_iterations) = 1320

[hjh0119]

Hope your question has been resolved. If you have any further issues, feel free to reopen the issue.

[wwwjjjj]

Hope your question has been resolved. If you have any further issues, feel free to reopen the issue.

Hello, thank you for your reply. The issue has been resolved.