GRPO 训练算力要求

[metro8116]

您好，请问要实现同等参数配置的grpo最低算力要求是什么？对显卡显存的选用有什么建议吗？

[jonasjunzhang]

Hi, thanks for your interest.

For our released GRPO setup, we trained with 8 NVIDIA H20 GPUs (96 GB each).

If GPU memory is tight, the first thing to try is changing the DeepSpeed ZeRO stage. The current GRPO script uses scripts/zero1.json by default, and you can switch --deepspeed_config to scripts/zero3.json, which should reduce per-GPU memory usage noticeably.

Other parameters you can adjust are:

• --per_device_train_batch_size: keep this at 1 (the default) if memory is limited.

Hope this helps.