需要的GPU hours

[XiangqianHuang]

您好，我想问下，如果是Reinforcement Fine-Tuning，你们在什么GPU上训练了多少小时，我担心我的显卡配置不够导致训练时间缓慢，期待您的回复。

[Luciennnnnnn]

Hi, thanks for your interest! For OmniGen2-EditScore7B-v1.1, training 700 steps takes about 20 hours on 32 H100 GPUs, with the reward server running on another 32 H100 GPUs.

However, you can use fewer GPUs by setting up a more efficient reward model (e.g., EditScore-7B or EditScore-7B-avg@4). The number of GPUs for the reward server is flexible — fewer GPUs will work too, although training may be slower. In addition, many RL training parameters can be adjusted, such as batch size and rollouts.

If you plan to RTF Flux Kontext, it should be faster — since Flux Kontext has undergone CFG distillation, which significantly reduces sampling time during RL training.

嗨，感谢你的关注！对于 OmniGen2-EditScore7B-v1.1，在 32 块 H100 GPU 上训练 700 步大约需要 20 小时，奖励服务器则运行在另外 32 块 H100 GPU 上。
不过，你也可以通过使用更高效的奖励模型（例如 EditScore-7B 或 EditScore-7B-avg@4）来使用更少的 GPU。奖励服务器的 GPU 数量也很灵活——少一些 GPU 也可以，只是训练可能会慢一些。此外，RL 训练的很多参数都是可调的，比如 batch size 和 rollout 步数。
如果你计划 RTF Flux Kontext，速度会更快——因为 Flux Kontext 已经过 CFG 蒸馏，在 RL 训练中可以显著减少采样时间。