problem about batch_size and gradient_accumulation_steps

[ShowLo]

In the paper, it's show that MuseTalk training was performed on 2 NVIDIA H20 GPUs, and the Unet model was initially trained with L1 and perceptual losses for 200,000 steps. However, the paper doesn't specify the batch_size and gradient_accumulation_steps, which impact training speed. Could you provide the specific numbers used?