Fyeward

> Try set smaller gradient_accumulation_steps in default_config.yaml. In GRPOTrainer cases, the memory of GPU will get significant larger with a larger gradient_accumulation_steps. The low use ratio of GPU is also...

> > > Try set smaller gradient_accumulation_steps in default_config.yaml. In GRPOTrainer cases, the memory of GPU will get significant larger with a larger gradient_accumulation_steps. The low use ratio of GPU...

> > > > > Try set smaller gradient_accumulation_steps in default_config.yaml. In GRPOTrainer cases, the memory of GPU will get significant larger with a larger gradient_accumulation_steps. The low use ratio...

> For the training yes, but not for the generation. The generation is done once once the full effective batch. Thank you. I noticed that, according to the improvement in...